Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a novel artificial intelligence model inspired by neural oscillations in the brain, with the goal of significantly advancing how machine learning algorithms handle long sequences of data.AI often struggles with analyzing complex information that unfolds over long periods of time, such as climate trends, biological signals, or financial data. One new type of AI model, called "state-space models," has been designed specifically to understand these sequential patterns more effectively. However, existing state-space models often face challenges — they can become unstable or require a significant amount of computational resources when processing long data sequences.To address these issues, CSAIL researchers T. Konstantin Rusch and Daniela Rus have developed what they call “linear oscillatory state-space models” (LinOSS), which leverage principles of forced harmonic oscillators — a concept deeply rooted in physics and observed in biological neural networks. This approach provides stable, expressive, and computationally efficient predictions without overly restrictive conditions on the model parameters."Our goal was to capture the stability and efficiency seen in biological neural systems and translate these principles into a machine learning framework," explains Rusch. "With LinOSS, we can now reliably learn long-range interactions, even in sequences spanning hundreds of thousands of data points or more."The LinOSS model is unique in ensuring stable prediction by requiring far less restrictive design choices than previous methods. Moreover, the researchers rigorously proved the model’s universal approximation capability, meaning it can approximate any continuous, causal function relating input and output sequences.Empirical testing demonstrated that LinOSS consistently outperformed existing state-of-the-art models across various demanding sequence classification and forecasting tasks. Notably, LinOSS outperformed the widely-used Mamba model by nearly two times in tasks involving sequences of extreme length.Recognized for its significance, the research was selected for an oral presentation at ICLR 2025 — an honor awarded to only the top 1 percent of submissions. The MIT researchers anticipate that the LinOSS model could significantly impact any fields that would benefit from accurate and efficient long-horizon forecasting and classification, including health-care analytics, climate science, autonomous driving, and financial forecasting."This work exemplifies how mathematical rigor can lead to performance breakthroughs and broad applications," Rus says. "With LinOSS, we’re providing the scientific community with a powerful tool for understanding and predicting complex systems, bridging the gap between biological inspiration and computational innovation."The team imagines that the emergence of a new paradigm like LinOSS will be of interest to machine learning practitioners to build upon. Looking ahead, the researchers plan to apply their model to an even wider range of different data modalities. Moreover, they suggest that LinOSS could provide valuable insights into neuroscience, potentially deepening our understanding of the brain itself. Their work was supported by the Swiss National Science Foundation, the Schmidt AI2050 program, and the U.S. Department of the Air Force Artificial Intelligence Accelerator.

“Linear oscillatory state-space models” leverage principles of forced harmonic oscillators — a concept deeply rooted in physics and observed in biological neural networks. This can improve how we predict complex information like climate trends or financial data. Image: Pixabay

New type of "state-space model" leverages principles of harmonic oscillators.

Novel AI model inspired by neural dynamics from the brain

﻿This article is based on a conversation with Fabio Rosa, CEO of Tago.io.Pilots are not to prove the technology, it's to prove your business caseIn today’s IoT landscape, the technology has proven itself. Sensors are accurate. Devices are reliable. Connectivity is everywhere; from 4G and 5G to LoRaWAN and non-terrestrial networks. And IoT platforms are powerful, flexible, and customizable.Yet, despite this maturity, too many IoT projects don't make it past the pilot phase.Why?Because companies often treat pilots as technical exercises, instead of what they truly are: business case validations.<h2>5 questions to ask yourself during the pilot phase</h2>A real pilot should force you, and your team, to face some hard truths. Here are five critical questions to ask yourself:<ol><li>What is the minimum required scale for this project to succeed? Companies often over-estimate the scale and speed of hardware rollouts. Take an honest look at the numbers and define the required scale of the project to make the numbers work. </li><li>If I would get free access to hardware / connectivity / IoT platforms, would the business case look significantly different? There’s an obsession with costs, leading to tough negotiations to scrape off cents per device. If these negotiations have a significant impact on the business case of your IoT project, something is awfully off.  </li><li>Am I ready to change the roles and responsibilities of my personnel if the pilot succeeds?  IoT solutions streamline operations and automate tasks previously handled manually. If your project delivers the expected efficiencies, it will impact job roles and performance metrics. Are you ready to adjust how you evaluate, reward, and organize your team? Are you truly ready for the consequences of success? </li><li>What transformation does my company need to go through when implementing this IoT project? New processes, new skills, new departments. A real pilot highlights the organizational changes you must be willing to make, and the likelihood of these changes to succeed. The metrics used to define the success of projects and performance need to be realigned. </li><li>What additional resources do I need for installing and maintaining the solution?  Every connected device requires maintenance, support, and updates. Factor those into your business case at this phase, not during the production phase. Leaders and managers should be prepared to run a marathon, not a sprint. Therefore, additional resources and budget are necessary to achieve the long-term goals.</li></ol><h2>Focus where it matters</h2>A successful pilot is not one where everything "works" from a technical perspective. A successful pilot is one that proves your business case, that shows you can scale economically, create real value, and transform your operations sustainably.When you start treating pilots this way, your entire IoT strategy will change for the better. If you want a partner who understands this mindset, and who can help you build IoT projects that go beyond the pilot phase, talk to the team of <a href="https://tago.io/" target="_blank">TagoIO</a>.  <h2>About TagoIO</h2>TagoIO is an industry-leading IoT platform designed to empower businesses and individuals to connect, manage, and analyze the growing amounts of data IoT devices generate. The all-in-one platform streamlines data processing and visualization by eliminating complexity, saving businesses time, and freeing up resources so they can innovate their products and services.The platform is trusted by thousands of businesses that customize and deploy as the backbone of their solutions to connect IoT devices to hundreds of thousands of users worldwide. TagoIO has become a go-to solution for many IoT applications across many verticals, such as Agriculture, Buildings, Energy, Industrial, and Logistics.From its headquarters in Raleigh, NC, USA, TagoIO supports domestic and international businesses in more than 130 countries to build the best IoT solution. For more information, visit <a href="http://www.tago.io" target="_blank">www.tago.io</a>

IoT projects do not fail because of the technology, they fail because of bad business cases.

5 questions to ask yourself during the pilot phase

If there’s one thing that characterizes driving in any major city, it’s the constant stop-and-go as traffic lights change and as cars and trucks merge and separate and turn and park. This constant stopping and starting is extremely inefficient, driving up the amount of pollution, including greenhouse gases, that gets emitted per mile of driving. One approach to counter this is known as eco-driving, which can be installed as a control system in autonomous vehicles to improve their efficiency.How much of a difference could that make? Would the impact of such systems in reducing emissions be worth the investment in the technology? Addressing such questions is one of a broad category of optimization problems that have been difficult for researchers to address, and it has been difficult to test the solutions they come up with. These are problems that involve many different agents, such as the many different kinds of vehicles in a city, and different factors that influence their emissions, including speed, weather, road conditions, and traffic light timing.“We got interested a few years ago in the question: Is there something that automated vehicles could do here in terms of mitigating emissions?” says Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor in the Department of Civil and Environmental Engineering and the Institute for Data, Systems, and Society (IDSS) at MIT, and a principal investigator in the Laboratory for Information and Decision Systems. “Is it a drop in the bucket, or is it something to think about?,” she wondered.To address such a question involving so many components, the first requirement is to gather all available data about the system, from many sources. One is the layout of the network’s topology, Wu says, in this case a map of all the intersections in each city. Then there are U.S. Geological Survey data showing the elevations, to determine the grade of the roads. There are also data on temperature and humidity, data on the mix of vehicle types and ages, and on the mix of fuel types.Eco-driving involves making small adjustments to minimize unnecessary fuel consumption. For example, as cars approach a traffic light that has turned red, “there’s no point in me driving as fast as possible to the red light,” she says. By just coasting, “I am not burning gas or electricity in the meantime.” If one car, such as an automated vehicle, slows down at the approach to an intersection, then the conventional, non-automated cars behind it will also be forced to slow down, so the impact of such efficient driving can extend far beyond just the car that is doing it.That’s the basic idea behind eco-driving, Wu says. But to figure out the impact of such measures, “these are challenging optimization problems” involving many different factors and parameters, “so there is a wave of interest right now in how to solve hard control problems using AI.” The new benchmark system that Wu and her collaborators developed based on urban eco-driving, which they call “IntersectionZoo,” is intended to help address part of that need. The benchmark was described in detail in a <a href="https://openreview.net/pdf?id=XoulHHQGFi">paper</a> presented at the 2025 International Conference on Learning Representation in Singapore.Looking at approaches that have been used to address such complex problems, Wu says an important category of methods is multi-agent deep reinforcement learning (DRL), but a lack of adequate standard benchmarks to evaluate the results of such methods has hampered progress in the field.The new benchmark is intended to address an important issue that Wu and her team identified two years ago, which is that with most existing deep reinforcement learning algorithms, when trained for one specific situation (e.g., one particular intersection), the result does not remain relevant when even small modifications are made, such as adding a bike lane or changing the timing of a traffic light, even when they are allowed to train for the modified scenario.In fact, Wu points out, this problem of non-generalizability “is not unique to traffic,” she says. “It goes back down all the way to canonical tasks that the community uses to evaluate progress in algorithm design.” But because most such canonical tasks do not involve making modifications, “it’s hard to know if your algorithm is making progress on this kind of robustness issue, if we don’t evaluate for that.”While there are many benchmarks that are currently used to evaluate algorithmic progress in DRL, she says, “this eco-driving problem features a rich set of characteristics that are important in solving real-world problems, especially from the generalizability point of view, and that no other benchmark satisfies.” This is why the 1 million data-driven traffic scenarios in IntersectionZoo uniquely position it to advance the progress in DRL generalizability.  As a result, “this benchmark adds to the richness of ways to evaluate deep RL algorithms and progress.”And as for the initial question about city traffic, one focus of ongoing work will be applying this newly developed benchmarking tool to address the particular case of how much impact on emissions would come from implementing eco-driving in automated vehicles in a city, depending on what percentage of such vehicles are actually deployed.But Wu adds that “rather than making something that can deploy eco-driving at a city scale, the main goal of this study is to support the development of general-purpose deep reinforcement learning algorithms, that can be applied to this application, but also to all these other applications — autonomous driving, video games, security problems, robotics problems, warehousing, classical control problems.”Wu adds that “the project’s goal is to provide this as a tool for researchers, that’s openly available.” IntersectionZoo, and the documentation on how to use it, are freely available at <a href="https://github.com/mit-wu-lab/IntersectionZoo">GitHub</a>.Wu is joined on the paper by lead authors Vindula Jayawardana, a graduate student in MIT’s Department of Electrical Engineering and Computer Science (EECS); Baptiste Freydt, a graduate student from ETH Zurich; and co-authors Ao Qu, a graduate student in transportation; Cameron Hickert, an IDSS graduate student; and Zhongxia Yan PhD ’24.

“We got interested a few years ago in the question, is there something that automated vehicles could do here in terms of mitigating emissions,” says MIT Professor Cathy Wu. “Is it a drop in the bucket, or is it something to think about?” Image: iStock

"IntersectionZoo," a benchmarking tool, uses a real-world traffic problem to test progress in deep reinforcement learning algorithms.

New tool evaluates progress in reinforcement learning

Spectroscopy technology is critical for numerous industries, facilitating precise analysis of physical, chemical, and structural properties through interactions between matter and infrared light. Hamamatsu Photonics introduces the FTIR Engine C16511-01, a state-of-the-art near-infrared spectrometer designed for rapid and accurate data acquisition in demanding industrial environments. Leveraging advanced optical interferometry, Ethernet connectivity, and a compact form factor, this instrument enhances real-time monitoring capabilities in sectors such as pharmaceuticals, food processing, and biotechnology. By addressing previous technological limitations, Hamamatsu’s innovation provides greater accuracy, reliability, and integration flexibility, significantly advancing the application of industrial spectroscopy.<h2>Introduction</h2>Optical spectroscopy is a technique that analyzes how light interacts with matter to determine its physical, chemical, or structural properties using wavelengths in the ultraviolet (UV), visible, or infrared spectrum. Increasingly, spectroscopy is essential in sectors such as pharmaceutical manufacturing, food processing, and biotechnology, where rapid and accurate real-time data acquisition is crucial for quality control and operational efficiency [1]. Responding to these industry demands, Hamamatsu Photonics has released the <a href="https://www.hamamatsu.com/us/en/product/optical-sensors/spectrometers/ftir-engine/C16511-01.html">FTIR Engine C16511-01</a>. It is engineered to overcome limitations of traditional spectrometers, including slow data processing, limited integration flexibility, and inadequate communication interfaces [2].<h3>Current Challenges in Optical Spectroscopy</h3>Traditionally, industrial spectroscopy faced substantial challenges, primarily involving complex instrumentation and limited data transfer speeds. Bulky equipment often cannot fit easily into crowded factory environments. Many traditional systems rely on USB connections, limiting data speed and reliability, which complicates rapid decision-making in industrial production [2]. These industrial diffractive spectrometers also face a critical trade-off between spectral resolution and measurement speed. FTIR spectrometers, however, overcome this compromise, delivering high spectral resolution simultaneously with high-speed data collection [2].Moreover, environmental factors such as temperature variations, vibrations from machinery, and electromagnetic interference can affect optical components. These factors make it challenging to maintain consistently high wavelength accuracy. Precision alignment of mirrors and detectors also poses difficulties over extended operational periods, especially in environments where frequent calibration may be impractical. Uncalibrated measurements become unstable and degrade the quality of collected data, leading to less confident quality control outcomes.Another significant issue is the variability in sample conditions and analyte concentrations, which can demand real-time monitoring. Bulky or sensitive spectrometers may have limited tolerance for environmental fluctuations, necessitating controlled laboratory-like conditions that are not always feasible in process lines [3]. Consequently, industrial facilities experience production slowdowns, higher operational costs, and elevated risk of noncompliant products. When high wavelength accuracy and stable data acquisition are not ensured, results can become inconsistent or erroneous, necessitating rework or scrapped batches, directly impacting a company’s competitiveness [4].In short, the combination of physical hardware constraints, suboptimal communication interfaces, environmental stressors, and calibration complexities can all impede successful spectroscopic analysis [5]. Addressing these limitations is essential for safeguarding product integrity, optimizing resource usage, and enabling effective real-time decision-making in modern industrial ecosystems [1]. <h2>Introducing the FTIR Engine C16511-01</h2>The <a href="https://www.hamamatsu.com/us/en/product/optical-sensors/spectrometers/ftir-engine/C16511-01.html">FTIR Engine C16511-01</a> from Hamamatsu integrates advanced Fourier Transform Infrared (FTIR) technology within a compact, rugged enclosure optimized for industrial applications [1]. By utilizing Ethernet (GigE Vision) connectivity, this new instrument ensures significantly enhanced communication speed and reliability compared to traditional USB-connected spectrometers [1,2].<h3><iframe id="680910fd4f6c9900d9178b4a" width="250" class="specs-iframe" height="440" src="https://wevolver.com/iframe/specs/ftir-engine-ft-nir-spectrometer?iframe=true" frameborder="0" scrolling="no"></iframe>Advanced Optical Interferometer Technology</h3>Central to the FTIR Engine C16511-01 is its Michelson optical interferometer, which includes a built-in light input section, beam splitter, fixed mirror, movable mirror (ϕ3 mm), and a high-performance InGaAs photodetector. This configuration precisely captures variations in light intensity signals caused by the position of the movable mirror. These signals are processed using a Fourier transformation, producing highly accurate optical spectra [1].Additionally, the system features an integrated vertical cavity surface-emitting laser (VCSEL). A VSCEL is a specialized semiconductor laser that emits light perpendicular to the wafer surface, providing continuous, precise monitoring of the movable mirror’s position. This distinctive laser design ensures superior wavelength accuracy and thermal stability by producing a stable and narrow emission profile, which is critical for sensitive industrial processes requiring rigorous quality control. By delivering a consistent reference signal, the VCSEL minimizes measurement drift and enhances the overall precision of Fourier transform calculations, thus maintaining reliable spectral data throughout demanding industrial operations [1].<h2>Detailed Technical Capabilities</h2><h3>High Signal-to-Noise Performance</h3>With a signal-to-noise ratio exceeding 10,000, the device can reliably detect and quantify subtle variations in spectral data. This sensitivity ensures accurate measurements even in environments with inherent background noise, enabling precise detection of low-concentration analytes critical for many industrial processes [1].<h3>High-Speed Data Acquisition</h3>The FTIR Engine C16511-01 records spectra at a rate of 275 frames per second. This capability is crucial for capturing dynamic processes in real-time, essential for rapidly evolving reactions or high-throughput industrial environments where timely data is critical [1].<h3>Ethernet Connectivity</h3>The Ethernet (GigE Vision) interface provides rapid and reliable data transfer, fully compatible with Windows and Linux operating systems, offering easy integration with OEM equipment. It ensures robust, continuous performance even in electromagnetically noisy industrial environments [1,2].<h3>Compact Form Factor</h3>The compact design, measuring 68 × 124 × 66 mm and weighing approximately 580 g, facilitates integration into constrained or mobile industrial settings without sacrificing durability. This form factor enables installation in various operational configurations, including robotic or automated systems, where space efficiency is paramount [1].<h3>High Wavelength Accuracy</h3>The built-in VCSEL laser provides precise positional feedback of the movable mirror, maintaining consistent wavelength accuracy. This capability ensures reliable quantitative and qualitative analyses across multiple measurements, which is critical in regulatory-sensitive applications such as pharmaceuticals and food safety [1].<h3>Broad Spectral Range</h3>Covering wavelengths from 1100 nm to 2500 nm, the FTIR Engine enables comprehensive analysis across a wide variety of chemical compounds. This extensive spectral coverage supports diverse industrial applications, facilitating detailed molecular characterization essential for quality control and research [1].<h3>Versatile Industrial Applications</h3>The FTIR Engine C16511-01 is a versatile analytical tool supporting various industries. In pharmaceuticals, it enables rapid, non-destructive inspection of medicinal products, ensuring accurate identification and quantification of active pharmaceutical ingredients (APIs) and excipients. It also supports polymorphism studies essential for drug solubility, stability, and bioavailability, and quickly detects cross-contamination between production batches, preventing costly disruptions [5,8].Within food processing and agriculture, the engine rapidly authenticates ingredients, detects contaminants, and evaluates nutritional profiles, providing immediate quality assurance. Its non-destructive nature and minimal sample preparation streamline routine testing, ensuring consistency and safety in food production [6,9].Biotechnology industries utilize the FTIR Engine for real-time monitoring of biochemical reactions, gaining valuable insights into reaction dynamics, yield optimization, and purity control. It plays a vital role in downstream processing, verifying protein and enzyme purity, and preventing costly batch failures by identifying deviations early [7,10].Additionally, the FTIR Engine significantly enhances recycling processes by rapidly identifying various polymers such as polyethylene, polypropylene, and polystyrene. Automated inline systems leverage its capabilities for efficient sorting of multi-layer plastics, improving recycling accuracy and supporting sustainable waste management [5,11].<h2>User Benefits and Integration</h2>The FTIR Engine C16511-01 provides significant operational advantages, including high precision and rapid measurements for improved decision-making and quality assurance. Its robust Ethernet connectivity ensures reliable, high-speed data transfer, reducing downtime and maintenance requirements. The compact design simplifies integration across diverse industrial setups, enhancing flexibility.To further support customers, Hamamatsu’s FTIR Engine comes with intuitive evaluation software. This software facilitates streamlined data acquisition, real-time configuration, and spectral visualization. It logs extensive spectral data and supports automated batch processing. Additionally, detailed DLL specifications allow integrators to develop custom software modules using languages such as C++ or Python, tailored to specific analytical needs. These APIs seamlessly integrate with existing SCADA systems and synchronize with third-party sensors or actuators, significantly improving overall process control and operational efficiency [1].<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzQ1NDE3NDYzNjYyLTExQTd3XzZBTHlXN0VzNUJIUTR6b0JsOVFiUnF2QjlhLmpwZyIsImVkaXRzIjp7InJlc2l6ZSI6eyJ3aWR0aCI6OTUwLCJmaXQiOiJjb3ZlciJ9fX0=" style="width:833px" class="fr-fic fr-dib" />Connection diagram for transmission measurement.<h2 id="isPasted">Additional Options and Accessories</h2>Hamamatsu offers several accessories that further improve reliability and integration flexibility. Vibration isolation tables ensure measurement stability, specialized optical fiber cables minimize signal distortion, and flexible power options (12 V or 24 V) accommodate diverse industrial requirements. The FTIR Engine is also compatible with the L16462-01, a fiber-coupled diffused reflection light source designed for near-infrared applications. This light source enhances signal collection for diffuse reflectance measurements, improving data quality in in-line or remote monitoring applications. Together, these options ensure consistent data quality in demanding operational conditions [1].<h3>Enhance Your Real-Time Monitoring Capabilities</h3>The FTIR Engine C16511-01 offers high-speed, high-accuracy spectral data acquisition in a compact, industrial-ready design. With advanced optical interferometry, integrated VCSEL technology, and robust Ethernet connectivity, it supports reliable in-line analysis across a range of demanding applications—from pharmaceutical manufacturing to polymer sorting.<a href="https://www.hamamatsu.com/eu/en/product/optical-sensors/spectrometers/ftir-engine/C16511-01.html">Learn more about the FTIR Engine C16511-01 and how it can support your analytical and process control requirements.</a><h2>References</h2>[1] Hamamatsu FTIR Engine C16511-01 Technical Documentation, Hamamatsu Photonics[2] Hamamatsu FTIR Engine C15511-01 Technical Documentation (Previous Version), Hamamatsu Photonics[3] Mitchell, S. (2020). "Calibration complexities in near-infrared reflectance technology for high-throughput lines." Journal of Applied Spectroscopy, 47(1), 11–15.[4] Smith, J. &amp; Wilhelm, K. (2021). "Transmission Bottlenecks in High-Volume Spectral Data Systems." Industrial Engineering Journal, 58(2), 89–97.[5] FTIR Spectroscopy Testing and Analysis, Contract Laboratory: https://outsource.contractlaboratory.com/ftir-spectroscopy-testing-and-analysis[6] Agilent Technologies, FTIR Applications in Food Testing: https://www.agilent.com/library/applications/5991-1246EN_AppNote_MM_FoodIngredients.pdf[7] Mettler Toledo, FTIR Spectroscopy for Bioprocess Monitoring:<a href="https://www.mt.com/us/en/home/applications/L1_AutoChem_Applications/ftir-spectroscopy.html"> https://www.mt.com/us/en/home/applications/L1_AutoChem_Applications/ftir-spectroscopy.html</a>[8] Huang, T., &amp; Carter, S. (2022). "Advancements in FTIR-Based Contamination Detection in Pharmaceutical Manufacturing." Journal of Pharmaceutical Analysis, 34(3), 211–219.[9] Rossi, L., &amp; Martins, J. (2021). "Nutrient Profiling Using FTIR Spectroscopy in Food Production." Food Quality and Safety Journal, 17(4), 98–104.[10] Pérez, M. &amp; Clarkson, D. (2023). "Real-Time Downstream Process Monitoring with FTIR in Biotechnology." Bioprocess Engineering Quarterly, 45(2), 67–72.[11] Tanaka, Y. &amp; Basu, R. (2021). "Automated Inline FTIR Analysis for Multi-Layer Polymer Recycling." Journal of Sustainable Materials Engineering, 22(1), 13–21.

Hamamatsu Photonics introduces the FTIR Engine C-, a state-of-the-art near-infrared spectrometer designed for rapid and accurate data acquisition in demanding industrial environments.

Compact, High-Speed Spectroscopy for Real-Time Applications: The FTIR Engine C-

Data privacy comes with a cost. There are security techniques that protect sensitive user data, like customer addresses, from attackers who may attempt to extract them from AI models — but they often make those models less accurate.MIT researchers recently developed a framework, based on a new privacy metric called PAC Privacy, that could maintain the performance of an AI model while ensuring sensitive data, such as medical images or financial records, remain safe from attackers. Now, they’ve taken this work a step further by making their technique more computationally efficient, improving the tradeoff between accuracy and privacy, and creating a formal template that can be used to privatize virtually any algorithm without needing access to that algorithm’s inner workings.The team utilized their new version of PAC Privacy to privatize several classic algorithms for data analysis and machine-learning tasks.They also demonstrated that more “stable” algorithms are easier to privatize with their method. A stable algorithm’s predictions remain consistent even when its training data are slightly modified. Greater stability helps an algorithm make more accurate predictions on previously unseen data.The researchers say the increased efficiency of the new PAC Privacy framework, and the four-step template one can follow to implement it, would make the technique easier to deploy in real-world situations.“We tend to consider robustness and privacy as unrelated to, or perhaps even in conflict with, constructing a high-performance algorithm. First, we make a working algorithm, then we make it robust, and then private. We’ve shown that is not always the right framing. If you make your algorithm perform better in a variety of settings, you can essentially get privacy for free,” says Mayuri Sridhar, an MIT graduate student and lead author of a <a href="https://eprint.iacr.org/2024/718.pdf" target="_blank">paper on this privacy framework</a>.She is joined in the paper by Hanshen Xiao PhD ’24, who will start as an assistant professor at Purdue University in the fall; and senior author Srini Devadas, the Edwin Sibley Webster Professor of Electrical Engineering at MIT. The research will be presented at the IEEE Symposium on Security and Privacy.<h2>Estimating noise</h2>To protect sensitive data that were used to train an AI model, engineers often add noise, or generic randomness, to the model so it becomes harder for an adversary to guess the original training data. This noise reduces a model’s accuracy, so the less noise one can add, the better.PAC Privacy automatically estimates the smallest amount of noise one needs to add to an algorithm to achieve a desired level of privacy.The original PAC Privacy algorithm runs a user’s AI model many times on different samples of a dataset. It measures the variance as well as correlations among these many outputs and uses this information to estimate how much noise needs to be added to protect the data.This new variant of PAC Privacy works the same way but does not need to represent the entire matrix of data correlations across the outputs; it just needs the output variances.“Because the thing you are estimating is much, much smaller than the entire covariance matrix, you can do it much, much faster,” Sridhar explains. This means that one can scale up to much larger datasets.Adding noise can hurt the utility of the results, and it is important to minimize utility loss. Due to computational cost, the original PAC Privacy algorithm was limited to adding isotropic noise, which is added uniformly in all directions. Because the new variant estimates anisotropic noise, which is tailored to specific characteristics of the training data, a user could add less overall noise to achieve the same level of privacy, boosting the accuracy of the privatized algorithm.<h2>Privacy and stability</h2>As she studied PAC Privacy, Sridhar hypothesized that more stable algorithms would be easier to privatize with this technique. She used the more efficient variant of PAC Privacy to test this theory on several classical algorithms.Algorithms that are more stable have less variance in their outputs when their training data change slightly. PAC Privacy breaks a dataset into chunks, runs the algorithm on each chunk of data, and measures the variance among outputs. The greater the variance, the more noise must be added to privatize the algorithm.Employing stability techniques to decrease the variance in an algorithm’s outputs would also reduce the amount of noise that needs to be added to privatize it, she explains.“In the best cases, we can get these win-win scenarios,” she says.The team showed that these privacy guarantees remained strong despite the algorithm they tested, and that the new variant of PAC Privacy required an order of magnitude fewer trials to estimate the noise. They also tested the method in attack simulations, demonstrating that its privacy guarantees could withstand state-of-the-art attacks.“We want to explore how algorithms could be co-designed with PAC Privacy, so the algorithm is more stable, secure, and robust from the beginning,” Devadas says. The researchers also want to test their method with more complex algorithms and further explore the privacy-utility tradeoff.“The question now is: When do these win-win situations happen, and how can we make them happen more often?” Sridhar says.“I think the key advantage PAC Privacy has in this setting over other privacy definitions is that it is a black box — you don’t need to manually analyze each individual query to privatize the results. It can be done completely automatically. We are actively building a PAC-enabled database by extending existing SQL engines to support practical, automated, and efficient private data analytics,” says Xiangyao Yu, an assistant professor in the computer sciences department at the University of Wisconsin at Madison, who was not involved with this study.This research is supported, in part, by Cisco Systems, Capital One, the U.S. Department of Defense, and a MathWorks Fellowship.

MIT researchers enhanced a data privacy technique so it is more computationally efficient and increases the accuracy of the AI algorithms to which it is applied. Image: MIT News; iStock

The approach maintains an AI model’s accuracy while ensuring attackers can’t extract secret information.

New method efficiently safeguards sensitive AI training data

Researchers from TU/e have transmitted data between the TU/e campus and the High Tech Campus over a 4.6 km distance using infrared light. This happened at the astonishing rate of 5.7 terabits per second, the equivalent of streaming 1.9 million Netflix shows in HD simultaneously. This is the fastest wireless data transmission ever demonstrated over this big a distance in an urban setting.The <a href="https://arxiv.org/pdf/2503.21058" target="_blank" rel="noreferrer">record-breaking connection</a> was established between TU/e in the north and the High Tech Campus (HTC) in the south of Eindhoven using advanced optical antennas from Aircision, a spin-off of TNO that develops products for ultra-high capacity optical wireless systems. The company is located at the HTC.Their antennas transmit data through invisible infrared beams instead of cables or radio signals. This technique, known as free-space optical (FSO) communication, enables ultra-fast, interference-free data transmission.Infrared wireless communication combines the high data speeds known from optical fibers with the flexibility of wireless communication systems.Vincent van Vliet, PhD researcher“We need new ways to meet the increasing demand for fast and reliable connectivity,” says <a href="https://research.tue.nl/nl/persons/vincent-van-vliet" target="_blank" rel="noreferrer">Vincent van Vliet</a>, a TU/e PhD researcher involved in the project.<h3>Breaking boundaries in wireless speed </h3>“Infrared wireless communication combines the high data speeds known from optical fibers with the flexibility of wireless communication systems. By complementing existing wireless and fiber-optic technologies, we can build the densely interconnected networks required to bring high-speed data connectivity to every corner of the planet.”<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1744719540053-csm_BvOF+2023_0426_ABA+-+HCOT+lab_338cf0c9ed.jpg" style="width:100%px" class="fr-fic fr-dib" />Vincent van Vliet working in the laboratory. Photo: Bart van OverbeekeTo achieve this breakthrough, the team used the Reid Photonloop testbed. The TU/e has launched this permanent set-up to experiment with high-speed wireless communication. It uses cutting-edge technology that combines multiple wavelengths in a single transmission. This technique, commonly used in fiber optics, has now been successfully applied to transmit data with this high speed over this big a distance for the first time in urban wireless communication.“Because the transmitted infrared light is highly focused, an almost unlimited number of communication links can exist side-by-side without interference, allowing wireless network capacity growth at an unprecedented scale”, Van Vliet explains.<h3>A permanent test facility for next-gen networks  </h3>One end of the Reid Photonloop is located on the top of the Flux building of the TU/e campus, which houses the departments of <a href="https://www.tue.nl/en/our-university/departments/electrical-engineering" target="_blank">Electrical Engineering</a> and <a href="https://www.tue.nl/en/our-university/departments/applied-physics-and-science-education" target="_blank">Applied Physics and Science Education</a>. The other end of the test bed, 4.6 kilometers away on the other side of the city of Eindhoven, sits at the top of building 37 at the High Tech CampusThe findings were presented at the Optical Fiber Communications (OFC) Conference 2025 in San Francisco earlier this month.<a href="https://www.tue.nl/en/research/researchers/chigo-okonkwo" target="_blank">Chigo Okonkwo</a>, Associate Professor and head of TU/e’s High-Capacity Optical Transmission Lab, emphasizes the importance of the testbed. “This facility will allow us to refine high-speed wireless communication and optimize its reliability and availability in all weather conditions”, he says.<h3>Real-life deployments</h3><a href="https://www.aircision.com/" target="_blank" rel="noreferrer">Aircision</a> is already looking at how this technology can be used in real-world applications, such as wirelessly connecting new 5G/6G antennas for backhauling the existing network to bring high-speed internet access to areas where laying fiber is impractical or too costly.Luis Oliveira, co-founder of Aircision, is optimistic: “We are redefining how data is transmitted over the air. This record-breaking achievement proves that our technology is ready to make high-speed internet accessible to millions of people faster than ever before.”The Reid Photonloop testbed is named after the late Aircision co-founder John Reid, who was a driving force behind the development of the test bed but passed away before it was realized. The work is funded by the NWO TTW-Perspectief FREE P19-13, PhotonDelta National Growth Fund Programme on Photonics, and European Innovation Council Transition project CombTools (G.A. 101136978).<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1744719610593-csm_Bob_102+copy_974fef770f.jpg" style="width:100%px" class="fr-fic fr-dib" /><img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1744719617319-csm_Bob_17+copy+2_b45e73756c.jpg" style="width:100%px" class="fr-fic fr-dib" />

The optical antenna on the Flux roof overlooking Eindhoven, facing the direction of the High Tech Campus. Photo: Vincent van Vliet

TU/e researchers have sent the equivalent of 1.9 million Netflix series via infrared light from the roof of their building.

Invisible beams of light provide super-fast wireless data transfer

Today, almost everybody has heard of AI and millions around the world already use, or are exposed, to it - from ChatGPT writing our emails, to helping in medical diagnosis.At its base, AI uses algorithms – sets of mathematically rigorous instructions - that tell a computer how to perform a variety of advanced functions or transform facts into useful information. The Large Language Models (LLMs), that drive today’s increasingly powerful AI, are special kinds of algorithms that learn from massive, mostly centralized datasets.Yet, centralizing these huge datasets generates issues around security, privacy and the ownership of data – indeed the phrase ‘data is the new oil’ signifies that it has become a crucial resource, driving innovation and growth in today’s digital economy.To counter these concerns, an approach called federated learning is now revolutionizing AI. Contrary to training AI models on huge, centralized datasets, federated learning allows these models to learn across a network of decentralized devices (or servers), keeping the raw data at its source.<h2>Untrusting Data</h2>“Today’s AI trained with federated learning gathers data from all over the world – the internet, other large databases, hospitals, smart devices and so on. These systems are very effective but at the same time there’s a paradox. What makes them so effective also makes them very vulnerable to learning from ‘bad’ data,” explains Professor <a href="https://dcl.epfl.ch/rachid/">Rachid Guerraoui</a>, Head of the <a href="https://dcl.epfl.ch/site/">Distributed Computing Laboratory</a> (DCL) in the <a href="https://www.epfl.ch/schools/ic/">School of Computer and Communication Sciences</a>.Data can be bad for many reasons. Perhaps a lack of attention or human error means it is incorrectly entered into a database, maybe there are mistakes in the data to begin with, perhaps sensors or other instruments are broken or malfunctioning, incorrect or dangerous data may be recorded maliciously, etc. Sometimes, the data is good but the machine hosting it is hacked or bogus. In any case, if this data is used to train AI, it makes the systems less trustworthy and unsafe.“All this brings up one key question,” says Guerraoui, “can we build trustworthy AI systems without trusting any individual source of data?” After a <a href="https://dcl.epfl.ch/site/safeai">decade of theoretical work</a> dedicated to addressing this challenge, the professor and his team say the answer is yes! A <a href="https://link.springer.com/book/10.1007/978-981-97-0688-4" target="_blank" rel="noopener">recent book</a> summarizes their main findings.<h2>Trusting Datasets</h2>In collaboration with the <a href="https://www.inria.fr/en" target="_blank" rel="noopener">French National Institute for Research in Digital Science and Technology</a>, they are now putting their ideas to work. They have developed <a href="https://actu.epfl.ch/news/trustworthy-ai-without-trusted-data/byzfl.epfl.ch">ByzFL</a>, a library using the Python programming language that is designed to benchmark, and improve, federated learning models against adversarial threats, particular bad data.“We believe that the majority of data is good but how do we know which datasets we can’t trust?” asks Guerraoui. “Our ByzFL library tests whether a system is robust against priori unknown attacks and then makes that system more robust. More specifically, we give users software to emulate bad data for testing as well as including security filters to ensure robustness. The bad data is often distributed in a subtle way so that it’s not immediately visible.”ByzFL doesn’t isolate and locate good from bad data but uses robust aggregation schemes (e.g., median) to ignore extreme inputs. For example, if three sensors record a temperature of 6, 7 and 9 degrees but another records -20, it ruins an entire computation. The ByzFL software excludes the extremes so that the impact of the bad data is limited, while information is aggregated.<h2>Ensuring that next-generation AI works</h2>Artificial intelligence is expected to touch every part of our lives in the not too distant future. Guerraoui argues that today, most companies use very primitive forms of AI, for example streaming platforms recommending movies or AI assistants helping to write text. If someone doesn’t like the movie they are recommended or an email isn’t perfect it’s no big deal.Looking ahead, for any application that is mission critical, such as diagnosing cancer, driving a car or controlling an aeroplane, safe AI is essential. “The day that we really put Generative AI in hospitals, cars or transport infrastructure I think we will see that safety is problematic because of bad data,” Guerraoui says. “The biggest challenge right now is going from what I call an animal circus to the real world with something that we can trust. For critical applications, we are far from the point where we can stop worrying about safety. The goal of ByzFL is to help bridge this gap.”<h2>A role for Switzerland</h2>The professor worries that it may take some big accidents for the public and policy makers to understand that the AI created to date shouldn’t be used for medicine, transport or anything mission critical and that the development of a new generation of safe and robust AI is essential.“I think Switzerland can play a role here because we have a tradition of seriousness. We build things that work, we can use the guarantee of Swiss quality to demonstrate a certification system using this kind of software to show that AI really is safe without trusting any individual component,” he concluded.ByzFL was designed and developed by John Stephen, <a href="https://www.linkedin.com/in/geovani-rizk/" target="_blank" rel="noopener">Geovani Rizk</a>,<a href="https://www.linkedin.com/in/marc-gonzalez-vidal-786681201/" target="_blank" rel="noopener">Marc Gonzalez Vidal</a>,<a href="https://www.linkedin.com/in/rafael-pinot-61a477bb/" target="_blank" rel="noopener">Rafael Pinot</a>,<a href="https://dcl.epfl.ch/rachid/">Rachid Guerraoui</a> (all from EPFL) and <a href="https://www.linkedin.com/in/francois-taiani-445a1027/" target="_blank" rel="noopener">Francois Taiani</a> (from INRIA).Mehdi El Mhamdi, Julian Steiner, Peva Blanchard, Nirupam Gupta, Rafael Pinot, Youssef Allouah, Abdellah El Mrini, John Stephan, Sadegh Farhadkhani, Geovani Rizk, Arsany Guiguis, Georgios Damaskinos, Sebastien Rouault, Richeek Patra, Mahsa Taziki, Hoang Le Nguyen and Alexandre Maurer are all students and postdocs who have all worked with Professor Guerraoui in the challenge of trustworthy AI systems without trusted data.

EPFL researchers developed a ground-breaking new tool to help build safer AI.

Trustworthy AI without Trusted Data

The use of terahertz waves, which have shorter wavelengths and higher frequencies than radio waves, could enable faster data transmission, more precise medical imaging, and higher-resolution radar.But effectively generating terahertz waves using a semiconductor chip, which is essential for incorporation into electronic devices, is notoriously difficult.Many current techniques can’t generate waves with enough radiating power for useful applications unless they utilize bulky and expensive silicon lenses. Higher radiating power allows terahertz signals to travel farther. Such lenses, which are often larger than the chip itself, make it hard to integrate the terahertz source into an electronic device.To overcome these limitations, MIT researchers developed a terahertz amplifier-multiplier system that achieves higher radiating power than existing devices without the need for silicon lenses.By affixing a thin, patterned sheet of material to the back of the chip and utilizing higher-power Intel transistors, the researchers produced a more efficient, yet scalable, chip-based terahertz wave generator.This compact chip could be used to make terahertz arrays for applications like improved security scanners for detecting hidden objects or environmental monitors for pinpointing airborne pollutants.“To take full advantage of a terahertz wave source, we need it to be scalable. A terahertz array might have hundreds of chips, and there is no place to put silicon lenses because the chips are combined with such high density. We need a different package, and here we’ve demonstrated a promising approach that can be used for scalable, low-cost terahertz arrays,” says Jinchen Wang, a graduate student in the Department of Electrical Engineering and Computer Science (EECS) and lead author of a paper on the terahertz&nbsp;radiator.He is joined on the paper by EECS graduate students Daniel Sheen and Xibi Chen; Steven F. Nagle, managing director of the T.J. Rodgers RLE Laboratory; and senior author Ruonan Han, an associate professor in EECS, who leads the Terahertz Integrated Electronics Group. The research will be presented at the IEEE International Solid-States Circuits Conference.<h2>Making waves</h2>Terahertz waves sit on the electromagnetic spectrum between radio waves and infrared light. Their higher frequencies enable them to carry more information per second than radio waves, while they can safely penetrate a wider range of materials than infrared light.One way to generate terahertz waves is with a CMOS chip-based amplifier-multiplier chain that increases the frequency of radio waves until they reach the terahertz range. To achieve the best performance, waves go through the silicon chip and are eventually emitted out the back into the open air.But a property known as the dielectric constant gets in the way of a smooth transmission.The dielectric constant influences how electromagnetic waves interact with a material. It affects the amount of radiation that is absorbed, reflected, or transmitted. Because the dielectric constant of silicon is much higher than that of air, most terahertz waves are reflected at the silicon-air boundary rather than being cleanly transmitted out the back.Since most signal strength is lost at this boundary, current approaches often use silicon lenses to boost the power of the remaining signal.&nbsp;The MIT researchers approached this problem differently.They drew on an electromechanical theory known as matching. With matching, they seek to equal out the dielectric constants of silicon and air, which will minimize the amount of signal that is reflected at the boundary.They accomplish this by sticking a thin sheet of material which has a dielectric constant between silicon and air to the back of the chip. With this matching sheet in place, most waves will be transmitted out the back rather than being reflected.<h2>A scalable approach</h2>They chose a low-cost, commercially available substrate material with a dielectric constant very close to what they needed for matching. To improve performance, they used a laser cutter to punch tiny holes into the sheet until its dielectric constant was exactly right.“Since the dielectric constant of air is 1, if you just cut some subwavelength holes in the sheet, it is equivalent to injecting some air, which lowers the overall dielectric constant of the matching sheet,” Wang explains.In addition, they designed their chip with special transistors developed by Intel that have a higher maximum frequency and breakdown voltage than traditional CMOS transistors.“These two things taken together, the more powerful transistors and the dielectric sheet, plus a few other small innovations, enabled us to outperform several other devices,” he says.Their chip generated terahertz signals with a peak radiation power of 11.1 decibel-milliwatts, the best among state-of-the-art techniques. Moreover, since the low-cost chip can be fabricated at scale, it could be integrated into real-world electronic devices more readily.One of the biggest challenges of developing a scalable chip was determining how to manage the power and temperature when generating terahertz waves.“Because the frequency and the power&nbsp;are so high, many of the standard ways to design a CMOS chip are not applicable here,” Wang says.The researchers also needed to devise a technique for installing the matching sheet that could be scaled up in a manufacturing facility.Moving forward, they want to demonstrate this scalability by fabricating a phased array of CMOS terahertz sources, enabling them to steer and focus a powerful terahertz beam with a low-cost, compact device.<hr>This research is supported, in part, by NASA’s Jet Propulsion Laboratory and Strategic University Research Partnerships Program, as well as the MIT Center for Integrated Circuits and Systems. The chip was fabricated through the Intel University Shuttle Program.

By affixing a thin, patterned sheet of material to the back of the chip, highlighted in the center and shown in the left-side micrograph, the researchers produced a more efficient, yet scalable, chip-based terahertz wave generator. Image: Courtesy of the researchers; MIT News

Researchers developed a scalable, low-cost device that can generate high-power terahertz waves on a chip, without bulky silicon lenses.

Chip-based system for terahertz waves could enable more efficient, sensitive electronics

The AI revolution has ushered in an era of exponential power and energy consumption. According to the <a href="https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers" rel="noopener" target="_blank">US Department of EnergyOpens in new window</a>, energy consumed by AI data centers could triple by 2028. Today, up to 40% of data center power use comes from cooling high-power chips—an astounding amount equivalent to the state of California’s entire electricity consumption.To combat this, Sheng Shen, professor of mechanical engineering at Carnegie Mellon University, has developed an innovative thermal interface material (TIM) that outperforms existing state-of-the-art solutions. His design, now published in <a href="https://www.nature.com/articles/s41467-025-56163-8" rel="noopener" target="_blank">Nature CommunicationsOpens in new window</a>, achieves ultra-low thermal resistance while increasing cooling efficiency via improved heat dissipation. It also proves to be highly reliable.<blockquote>While an immediate need is cooling data centers, this innovation can break through in industries that have been stuck using outdated thermal interface materials.<cite>Sheng Shen, Professor, Mechanical Engineering</cite></blockquote>“The material is like a bridge between the nano- and macroscopic worlds,” explained Zexiao Wang, Ph.D. candidate in Shen’s lab. “Because the nanoscale material can be created using macroscale approaches, we can see with our own eyes the impact of the material on the world.”Not only is Shen’s thermal interface material the best-performing on the market, it is also highly reliable. The team tested the material at extreme temperature ranges from -55 to 125 degree Celsius for more than 1,000 cycles, and the material showed no performance degradation.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1740445088809-0211-em-ai-data-center-cooling-breakthrough.png" style="width: 100%px;" class="fr-fic fr-dib">Novel thermal interface material designed by Sheng Shen.&nbsp;Source:&nbsp;College of Engineering“This material solves a lot of existing challenges, and is ready to be used today,” said Shen. “While an immediate need is focused on cooling data centers, the application for this innovation is extensive. It can break through in industries that have been stuck using outdated thermal interface materials. It can be used for pre-packaging, reworked when using non-adhesives, and enables thermal bonding of two substrates at room temperature.”“Oftentimes work at the nanoscale is foundational for a device that we might not see for decades,” said Qixian Wang, Ph.D. candidate. “It’s been exciting to see the real-world impact our material can have today because it is so easy to use.”“Our material will have great benefits to the field of AI computing,” said Rui Cheng, postdoc and innovation commercialization fellow of CMU and the lead author of the paper published in Nature Communications. “Beyond reducing energy consumption, we can make AI development more affordable, more renewable, and more reliable.”<iframe width="640" height="360" src="https://www.youtube.com/embed/BG4Ct4jYh3k?&amp;wmode=opaque&amp;enablejsapi=1&amp;origin=https://www.wevolver.com" frameborder="0" allowfullscreen="" class="fr-draggable" data-lf-form-tracking-inspected-kn9eq4roawkarlvp="true" data-lf-yt-playback-inspected-kn9eq4roawkarlvp="true" data-lf-vimeo-playback-inspected-kn9eq4roawkarlvp="true"></iframe>This research was supported by the National Science Foundation and ARPA-E (Advanced Research Projects - Energy). Collaborators included Tianyi Chen and Ana V. Garcia-Caraveo from the Oregon State University, and Navid Kazem and Loren Russell from Arieca, Inc.<hr>Sheng Shen and Rui Cheng are making this technology commercially available via their startup company,&nbsp;<a href="https://www.novolinc.com/" rel="noopener" target="_blank">NovoLINC, Inc.Opens in new window</a>

Novel thermal interface material cools chips better than existing state-of-the-art solutions.

Slashing AI data center cooling cost, GPU-CPU power use

Similar to how cryptocurrencies saw a shift from Web3.0 to decentralized finance, the introduction of Decentralized Physical Infrastructure Networks, or DePIN, seeks to decentralize control of technical resources such as computational power and wireless networks.&nbsp;Leveraging this technology, the platform <a href="https://www.aydo.ai/">AYDO</a> bridges the Internet of Things (IoT) and blockchain technologies by enabling data streaming from IoT devices to Decentralized Physical Infrastructure Networks (DePINs). Their platform allows users to connect third-party IoT devices and sensors to DePIN protocols, lowering the cost and complexity associated with data collection and integration. This enables private and decentralized data management directly on blockchain systems without the need for specialized hardware.AYDO’s solution benefits users by allowing them to stream and monetize their data through DePIN applications in a user-friendly app. It supports various IoT devices and allows participation in the DePIN ecosystem using existing hardware like Raspberry Pi, PCs, or servers. The platform supports multiple protocols, including Zigbee, Zwave, Modbus, and more, and is focused on expanding the adoption of IoT devices in decentralized applications.Wevolver sat down with AYDO CEO Alexander Markin to discuss the company's mission and impact on the IoT landscape.&nbsp;<hr>Wevolver: Alex, can you tell us a bit more about AYDO and the story behind it?Alex:&nbsp;Sure, Aydo was founded in 2023, building on a business model based on expertise gained from four years in smart homes and IoT connectivity. It enables users to connect third-party devices and share resources for environmental data projects, reducing hardware costs through blockchain. Users can participate without heavy investment, ensuring informed decisions on project quality.The current AYDO software solution is designed to support 90% of existing IoT devices, with compatibility established for widely used IoT communication protocols, including Zigbee, Z-Wave, Modbus, and RTSP. Support for two additional protocols, Matter and Thread, is planned for future implementation, allowing for coverage of the remaining 10% of IoT devices.<table style="width: 100%;"><tbody><tr><td style="width: 100%;">What is DePIN? Decentralized Physical Infrastructure Networks (DePINs) use blockchain and cryptoeconomic incentives to build and manage infrastructure. DePINs enable individuals and organizations to collectively develop and operate physical infrastructure like wireless networks, energy grids, and transportation systems. Participants earn financial rewards and ownership stakes through tokenized incentives, promoting efficiency and equity.</td></tr></tbody></table>Wevolver: Could you share the variety of data that can be exchanged on AYDO? How do you ensure that this data is collected and handled in a way that’s ethical and compliant with legal standards?Alex:&nbsp;The current focus is on environmental data, which is non-personal, to avoid privacy issues, with plans to work on personal data requiring a security infrastructure in the future.The platform does not guarantee the quality of projects but ensures that projects must provide collateral before requesting data, allowing users to make informed decisions about their participation. The platform does not guarantee the quality of projects but ensures that projects must provide collateral before requesting data, allowing users to make informed decisions about their participation. Users are encouraged to conduct their own due diligence by evaluating the project's team, legal entity, and business model, as well as analyzing tokenomics to assess potential value.Wevolver: Can you tell us some of the company's success stories or case studies?Alex:&nbsp;Current integration partners include <a href="http://375.ai">375.ai</a> for camera solutions, <a href="https://www.mysteriumvpn.com/">Mysterium VPN</a> for Internet sharing, and <a href="https://www.koii.network/">Koii network</a> for decentralized computation, with case studies available for highlighting in the article.Wevolver: Alex, can you tell us what the AYDO mission is?Alex: The mission is to bring deep data into mass adoption, allowing individuals to earn from their hardware and utilize it fully, with a vision of making data sharing as commonplace as ridesharing services.<table style="width: 100%;"><tbody><tr><td style="width: 100%;">Other examples of DePin applications:<ul><li>Decentralized Mapping Hivemapper incentivizes users to collect high-definition map data using 4K dashcams while driving. As of October 2024, Hivemapper has mapped over 330 million kilometers of roads, surpassing traditional mapping providers.</li><li>Decentralized Wireless Networks Helium builds a global IoT network by selling mini hotspots. Users earn HNT tokens by providing network coverage. By 2024, Helium had deployed over 1 million hotspots and expanded into mobile services.</li><li>High-Precision Positioning Services GEODNET uses rooftop satellite miners to enhance GPS accuracy, offering centimeter-level precision at a fraction of traditional GPS costs. The project has partnered with the U.S. Department of Agriculture, showcasing its potential in precision agriculture.</li></ul></td></tr></tbody></table>Wevolver:&nbsp;Could you describe the steps a new user would take to start using AYDO, from signing up to their first transaction?Alex: New users can register on the website, choose projects based on their data or computation needs, and participate without significant upfront costs, utilizing existing devices like Raspberry Pi or old laptops.<table style="width: 100%;"><tbody><tr><td style="width: 100%;">Getting started with AYDOTo set up AYDO from&nbsp;<a href="http://aydo.ai">Aydo.ai</a> on a Linux-based server (like Raspberry Pi or PC), follow these steps: <ol><li dir="ltr">Prepare the Hardware: Get a server/controller/hub that runs on Linux and a Zigbee Adapter.</li><li dir="ltr">Install Software: Execute the correct command to install the necessary software including NodeJS, Zigbee2Mqtt, and AYDO Core.</li><li dir="ltr">Open the AYDO app: After installation, connect your devices and start streaming data to get rewarded.</li></ol>These steps will help you deploy AYDO to develop Decentralized Physical Infrastructure Networks (DePIN) applications that integrate with IoT sensors. For more detailed information, visit the<a href="https://www.aydo.ai/"> AYDO </a>Guide. <img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM4NjQ5NDUzOTE0LTI1MDIwM19BWURPLTAxLmpwZyIsImVkaXRzIjp7InJlc2l6ZSI6eyJ3aWR0aCI6OTUwLCJmaXQiOiJjb3ZlciJ9fX0=" style="width: 100%px;" class="fr-fic fr-dib"></td></tr></tbody></table>Wevolver:&nbsp;And for engineers looking to learn more about Aydo, where can they get more information or connect with current platform users?Alex:&nbsp;We encourage anyone interested in learning more to follow our <a href="https://x.com/AYDO_AI">Twitter</a> for regular AMA sessions. We have a <a href="https://t.me/aydoai">Telegram</a> group for updates and a <a href="https://discord.com/invite/aydo">Discord</a> for questions. Currently, we are at the Open Beta Test stage so that you can get onboarded through our <a href="http://www.aydo.ai">website</a>.&nbsp;<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM4NjQ5NDk4MjY3LTI1MDEzMV9BWURPLTAxLmpwZyIsImVkaXRzIjp7InJlc2l6ZSI6eyJ3aWR0aCI6OTUwLCJmaXQiOiJjb3ZlciJ9fX0=" style="width: 100%px;" class="fr-fic fr-dib"><h3 dir="ltr" id="isPasted">Further reading:&nbsp;</h3><a href="https://osl.com/academy/article/what-is-depin-the-future-of-decentralized-physical-infrastructure-networks">https://osl.com/academy/article/what-is-depin-the-future-of-decentralized-physical-infrastructure-networks</a><a href="https://www.forbes.com/sites/digital-assets/2024/02/28/why-depins-may-be-the-next-big-narrative-in-web3/">https://www.forbes.com/sites/digital-assets/2024/02/28/why-depins-may-be-the-next-big-narrative-in-web3/</a>

AYDO CEO explains the strategic monetization and DePIN’s advancements in IoT. 

Decentralized Data: A Leadership Insight with AYDO's CEO, Alexander Markin

The neural network artificial intelligence models used in applications like medical image processing and speech recognition perform operations on hugely complex data structures that require an enormous amount of computation to process. This is one reason deep-learning models consume so much energy.To improve the efficiency of AI models, MIT researchers created an automated system that enables developers of deep learning algorithms to simultaneously take advantage of two types of data redundancy. This reduces the amount of computation, bandwidth, and memory storage needed for machine learning operations.Existing techniques for optimizing algorithms can be cumbersome and typically only allow developers to capitalize on either sparsity or symmetry — two different types of redundancy that exist in deep learning data structures.By enabling a developer to build an algorithm from scratch that takes advantage of both redundancies at once, the MIT researchers’ approach boosted the speed of computations by nearly 30 times in some experiments.Because the system utilizes a user-friendly programming language, it could optimize machine-learning algorithms for a wide range of applications. The system could also help scientists who are not experts in deep learning but want to improve the efficiency of AI algorithms they use to process data. In addition, the system could have applications in scientific computing.“For a long time, capturing these data redundancies has required a lot of implementation effort. Instead, a scientist can tell our system what they would like to compute in a more abstract way, without telling the system exactly how to compute it,” says Willow Ahrens, an MIT postdoc and co-author of a&nbsp;<a href="https://arxiv.org/pdf/2406.09266" target="_blank">paper on the system</a>, which will be presented at the International Symposium on Code Generation and Optimization.She is joined on the paper by lead author Radha Patel ’23, SM ’24 and senior author Saman Amarasinghe, a professor in the Department of Electrical Engineering and Computer Science (EECS) and a principal researcher in the Computer Science and Artificial Intelligence Laboratory (CSAIL).<h2>Cutting out computation</h2>In machine learning, data are often represented and manipulated as multidimensional arrays known as tensors. A tensor is like a matrix, which is a rectangular array of values arranged on two axes, rows and columns. But unlike a two-dimensional matrix, a tensor can have many dimensions, or axes, making tensors more difficult to manipulate.Deep-learning models perform operations on tensors using repeated matrix multiplication and addition — this process is how neural networks learn complex patterns in data. The sheer volume of calculations that must be performed on these multidimensional data structures requires an enormous amount of computation and energy.But because of the way data in tensors are arranged, engineers can often boost the speed of a neural network by cutting out redundant computations.For instance, if a tensor represents user review data from an e-commerce site, since not every user reviewed every product, most values in that tensor are likely zero. This type of data redundancy is called sparsity. A model can save time and computation by only storing and operating on non-zero values.In addition, sometimes a tensor is symmetric, which means the top half and bottom half of the data structure are equal. In this case, the model only needs to operate on one half, reducing the amount of computation. This type of data redundancy is called symmetry.“But when you try to capture both of these optimizations, the situation becomes quite complex,” Ahrens says.To simplify the process, she and her collaborators built a new compiler, which is a computer program that translates complex code into a simpler language that can be processed by a machine. Their compiler, called SySTeC, can optimize computations by automatically taking advantage of both sparsity and symmetry in tensors.They began the process of building SySTeC by identifying three key optimizations they can perform using symmetry.First, if the algorithm’s output tensor is symmetric, then it only needs to compute one half of it. Second, if the input tensor is symmetric, then algorithm only needs to read one half of it. Finally, if intermediate results of tensor operations are symmetric, the algorithm can skip redundant computations.<h2>Simultaneous optimizations</h2>To use SySTeC, a developer inputs their program and the system automatically optimizes their code for all three types of symmetry. Then the second phase of SySTeC performs additional transformations to only store non-zero data values, optimizing the program for sparsity.In the end, SySTeC generates ready-to-use code.“In this way, we get the benefits of both optimizations. And the interesting thing about symmetry is, as your tensor has more dimensions, you can get even more savings on computation,” Ahrens says.The researchers demonstrated speedups of nearly a factor of 30 with code generated automatically by SySTeC.Because the system is automated, it could be especially useful in situations where a scientist wants to process data using an algorithm they are writing from scratch.In the future, the researchers want to integrate SySTeC into existing sparse tensor compiler systems to create a seamless interface for users. In addition, they would like to use it to optimize code for more complicated programs.

The new compiler, called SySTeC, can optimize computations by automatically taking advantage of both sparsity and symmetry in tensors. Image: iStock

By automatically generating code that leverages two types of data redundancy, the system saves bandwidth, memory, and computation.

User-friendly system can help developers build more efficient simulations and AI models

Specialized hardware accelerators like Neural Processing Units (NPUs) and Tensor Processing Units (TPUs) have revolutionized artificial intelligence (AI) hardware by providing capabilities for accelerating machine learning tasks. Both are designed to optimize performance for tasks like machine learning, deep learning, and neural network computations, but they cater to distinct use cases and requirements.&nbsp;This article discusses the difference in architectures, performance, practical applications, and challenges of these powerful processing units, offering actionable insights for professionals.Recommended reading:&nbsp;<a href="https://www.wevolver.com/article/tpu-vs-gpu-in-ai-a-comprehensive-guide-to-their-roles-and-impact-on-artificial-intelligence" target="_blank">TPU vs GPU in AI: A Comprehensive Guide to Their Roles and Impact on Artificial Intelligence</a><h2 dir="ltr">Introduction</h2>The rapid advancement of AI has led to an increasing demand for specialized hardware capable of efficiently handling complex computations. NPUs and TPUs have emerged as pivotal technologies in this space, each tailored to optimize tasks involving complex machine learning algorithms in unique ways. NPUs, with their low-power design, enhance edge computing performance, making them indispensable for IoT and mobile devices. TPUs, on the other hand, excel in scaling deep learning workloads across massive datasets, driving breakthroughs in cloud-based AI.This article aims to provide professionals with a detailed comparison of these technologies, offering practical insights into their architectures, applications, and limitations. Understanding these distinctions can guide informed decision-making and optimize AI deployments for diverse requirements.<h2 dir="ltr">What Are Neural Processing Units (NPUs)?</h2>NPUs are specialized processors designed to optimize computations for neural networks, enabling efficient processing in real-time. With their ability to efficiently handle workloads like image recognition, voice processing, and natural language tasks, NPUs are an integral part of modern central processing units (CPUs) and System-on-Chip (SoC) designs. Companies like Huawei, Samsung, and Apple are leading the way in integrating NPUs into their devices to enhance AI capabilities without sacrificing power efficiency.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2MzI0NTQ0LWltYWdlMi5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="apple-neural-engine-devices">Apple's NPU, branded as the Neural Engine, is a key component of Apple's A-series and M-series chips. It is designed specifically for accelerating on-device machine learning and AI tasksRecommended reading:&nbsp;<a href="https://www.wevolver.com/article/npu-vs-gpu-understanding-the-key-differences-and-use-cases" target="_blank">NPU vs GPU: Understanding the Key Differences and Use Cases</a><h2 dir="ltr">What Are Tensor Processing Units (TPUs)?</h2>TPUs are hardware AI accelerator application-specific integrated circuits (ASICs) developed by Google to accelerate tensor-based computations, which are foundational to deep learning. Their design is optimized for mathematical operations like matrix-multiplication and parallel computing, making them ideal for tasks such as training large-scale neural networks and performing batch processing.Here is a block diagram showcasing the architecture of TPUs:<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2Mzc3NTczLWltYWdlNS5qcGciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="TPU-architecture-simplified-block-diagram">A simplified block diagram of a TPUThe TPU architecture is designed to optimize AI workloads by integrating specialized components for efficient computation, storage, and communication. At the core is the Matrix Multiply Unit, which performs rapid large-scale matrix operations critical for AI tasks. Data flows through the Unified Buffer, a local storage unit for managing activations, ensuring quick access and reducing dependency on external memory. The Control Unit orchestrates data movement and synchronizes operations between components. The PCIe Interface and Host Interface facilitate communication between the TPU and external systems, enabling smooth data transfer and coordination. The DRAM Ports provide access to external memory for managing large datasets that exceed the buffer's capacity. The Activation Pipeline processes intermediate results, applying transformations essential for refining AI computations. Additionally, the Accumulator stores partial computation results temporarily for aggregation and further processing. Finally, the Miscellaneous I/O manages input/output tasks, ensuring seamless integration with the broader system. This balanced architecture ensures high efficiency and performance for AI tasks in data centers.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2MzkwMzg5LWltYWdlNC5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="Google-TPU-v4">Google TPUs are primarily used in servers within data centers to accelerate large-scale AI workloads, such as training and deploying deep learning models.Recommended reading:&nbsp;<a href="https://www.wevolver.com/article/tensor-cores-vs-cuda-cores" target="_blank">Tensor Cores vs CUDA Cores: The Powerhouses of GPU Computing from Nvidia</a><h2 dir="ltr">Comparing the Performance of NPUs vs TPUs</h2>NPUs and TPUs exhibit distinct performance characteristics, making them suitable for different AI workloads. Following section compares both the AI chips based on different parameters.<h3 dir="ltr">Speed and Latency</h3>NPUs&nbsp;are designed for&nbsp;real-time, low-latency&nbsp;tasks, making them ideal for&nbsp;edge computing&nbsp;scenarios where immediate responses are required. For instance,&nbsp;NPUs&nbsp;enable instant image processing in&nbsp;smartphones&nbsp;and&nbsp;IoT devicesTPUs often excel in speed and scalability for large datasets, leveraging their specialized architecture for tensor computations. They are optimized for throughput rather than latency, making them suitable for batch data processing tasks like training massive deep learning models. In contrast, NPUs are designed for real-time, low-latency tasks on the edge.&nbsp;<h3 dir="ltr">Power and Energy Efficiency</h3>Power consumption is a critical factor for both NPUs and TPUs, but their optimization differs. NPUs provide better efficiency for edge computing and low-power applications, optimizing computing tasks with limited computational resources. &nbsp;This ensures extended battery life for small devices while performing complex AI tasks. TPUs, while efficient for their scale, are designed for environments like data centers where power availability is less constrained but efficiency remains essential for handling extensive computations.<h3 dir="ltr">Scalability and Flexibility</h3>NPUs&nbsp;are less scalable compared to&nbsp;TPUs, as they are tailored for specific edge&nbsp;AI&nbsp;applications. However, their&nbsp;flexibility&nbsp;in&nbsp;real-time&nbsp;processing tasks makes them invaluable for&nbsp;localized AI workloads.&nbsp;TPUs, on the other hand, are built for&nbsp;scalability, capable of handling&nbsp;large-scale&nbsp;AI&nbsp;models and&nbsp;datasets&nbsp;in&nbsp;cloud-based&nbsp;or&nbsp;distributed environments.<h2 dir="ltr">Architectural Differences</h2><h3 dir="ltr">Core Design</h3>NPUs&nbsp;are engineered to handle&nbsp;neural network activations&nbsp;and&nbsp;real-time&nbsp;matrix computations with specialized cores that focus on efficient, localized processing. These cores typically feature lightweight architectures optimized for low power consumption and latency-sensitive tasks.&nbsp;TPUs, in contrast, utilize systolic array architectures with a high density of computational cores designed for parallel matrix multiplications.<h3 dir="ltr">Programming Frameworks and Ecosystem</h3>NPUs are compatible with general-purpose AI frameworks such as TensorFlow, PyTorch, and ONNX, allowing developers to integrate AI tasks into various devices. As they are integrated within larger processors, they are more general-purpose and are compatible with a number of different computing tasks. TPUs, in contrast, are tightly integrated with Google’s TensorFlow ecosystem, making them an ideal choice for developers already working within Google’s AI infrastructure. The trade-off lies in flexibility versus specialization.<img src="https://images.wevolver.com/eyJidWNrZXQiOiJ3ZXZvbHZlci1wcm9qZWN0LWltYWdlcyIsImtleSI6ImZyb2FsYS8xNzM3NTY2NDEzOTA4LWltYWdlMS5wbmciLCJlZGl0cyI6eyJyZXNpemUiOnsid2lkdGgiOjk1MCwiZml0IjoiY292ZXIifX19" style="width: 100%px;" class="fr-fic fr-dib" alt="intel-core-ultra-gen-1-series-with-NPUs">Intel Core Ultra processors integrate NPUs to deliver efficient and enhanced performance for modern AI desktop applications.<h2 dir="ltr">Use Cases and&nbsp;Practical Applications</h2>NPUs and TPUs cater to distinct scenarios based on their specialized capabilities.NPUs find their strengths in edge computing, enabling AI tasks directly on devices like smartphones, cameras, and IoT sensors. Examples of&nbsp;NPU Applications&nbsp;include:<ul><li dir="ltr">Real-time object detection in edge devices, such as security cameras.</li><li dir="ltr">Voice recognition in smart assistants operating on low-power devices.</li><li dir="ltr">AI-powered healthcare devices for monitoring patient vitals and diagnosing conditions.</li><li dir="ltr">Industrial <a href="https://www.wevolver.com/article/building-automation-systems">automation systems</a> requiring real-time decision-making capabilities.</li></ul>TPUs are the backbone of AI operations on large amounts of data. Examples of&nbsp;TPU Applications&nbsp;include:<ul><li dir="ltr">Training large-scale machine learning models in cloud environments.</li><li dir="ltr">Batch processing for natural language understanding tasks in virtual assistants.</li><li dir="ltr">Optimizing recommendation systems in large-scale e-commerce platforms.</li><li dir="ltr">Enabling real-time analytics in large datasets for financial and scientific domains.</li></ul><h2 dir="ltr">Challenges and Limitations of NPUs and TPUs</h2><h3 dir="ltr">Engineering Challenges in Implementation</h3>Deploying NPUs and TPUs presents several technical challenges, including:<ul><li dir="ltr">Compatibility Issues: Existing software ecosystems, such as libraries and frameworks, often require extensive adaptation to fully leverage the capabilities of NPUs and TPUs.</li><li dir="ltr">Hardware Integration:&nbsp;Ensuring seamless integration with existing system architectures can be complex, particularly in legacy systems.</li><li dir="ltr">Optimization Challenges:&nbsp;AI models often need to be specifically optimized to achieve peak performance on these specialized processors.</li><li dir="ltr">Cost-Effectiveness:&nbsp;While powerful, the initial investment in NPU or TPU hardware can be prohibitive for smaller organizations.</li></ul><h3 dir="ltr">Limitations in Real-World Applications</h3>Despite their capabilities, NPUs and TPUs face limitations in specific environments:<ul><li dir="ltr">Latency-Sensitive Systems: TPUs, designed for high-throughput batch processing, may not be ideal for real-time applications that demand low latency.</li><li dir="ltr">Hybrid Requirements: Certain applications, such as autonomous vehicles, may require a combination of TPU and graphics processing unit (GPU) capabilities to balance precision and speed effectively.</li><li dir="ltr">Scalability Constraints:&nbsp;While TPUs scale well in cloud environments, NPUs are less adaptable to large-scale deployments due to their edge-specific design.</li></ul>Here is a table summarising the discussion on NPUs and TPUs:<div dir="ltr" align="left"><table><tbody><tr><td>Feature</td><td>NPUs</td><td>TPUs</td></tr><tr><td>Primary Use Case</td><td>Real-time AI tasks on edge devices (e.g., smartphones, laptops, IoT, robotics)</td><td>Large-scale AI workloads in data centers (e.g., training deep learning models)</td></tr><tr><td>Key Applications</td><td>Image recognition, natural language processing, voice recognition, augmented reality</td><td>Language models, natural language processing, large-scale image classification</td></tr><tr><td>Architectural Focus</td><td>Low-latency, power-efficient processing for real-time AI tasks</td><td>High-performance parallel processing and matrix multiplication for deep learning</td></tr><tr><td>Core Design</td><td>Lightweight architectures optimized for low power and real-time processing</td><td>Systolic array architectures optimized for massive parallel tensor operations</td></tr><tr><td>Scalability</td><td>Limited scalability, designed for specific edge AI tasks</td><td>High scalability for training large datasets and deploying AI models</td></tr><tr><td>Power Consumption</td><td>Optimized for energy efficiency in portable devices</td><td>Designed for efficiency in data centers with less strict power constraints</td></tr><tr><td>Programming Ecosystem</td><td>Supports general-purpose AI frameworks (e.g., TensorFlow, PyTorch, ONNX)</td><td>Tightly integrated with TensorFlow and Google Cloud services</td></tr><tr><td>Manufacturers</td><td>Companies like Apple (Neural Engine), Huawei, Samsung, and others</td><td>Google</td></tr><tr><td>Integration</td><td>Embedded in SoC designs for smartphones and IoT devices</td><td>Used as standalone accelerators in server environments</td></tr><tr><td>Key Strengths</td><td>Real-time processing, low latency, energy efficiency</td><td>Exceptional performance for large-scale AI training and inference</td></tr><tr><td>Key Limitations</td><td>Less suited for large-scale AI workloads</td><td>Higher power requirements, less suited for edge applications</td></tr></tbody></table></div><h2 dir="ltr">Conclusion</h2>NPUs and TPUs play complementary roles in the advancement of AI technology, each excelling in specific environments. While TPUs offer unparalleled performance for large-scale machine learning tasks in cloud environments, NPUs are perfect for real-time inference and energy-efficient edge applications.&nbsp;<h2 dir="ltr">Frequently Asked Questions</h2><ol><li dir="ltr">What Is the Primary Difference Between NPUs and TPUs? NPUs are specialized for low-power, real-time tasks, making them ideal for edge devices. TPUs are optimized for large-scale tensor computations, excelling in cloud-based machine learning.</li><li dir="ltr">What Are the Key Roles of NPUs and TPUs in AI? NPUs focus on real-time, low-power AI processing in edge devices like smartphones and IoT systems, enabling efficient inference for tasks such as image recognition and voice commands. TPUs specialize in high-performance, large-scale tensor computations, excelling in data centers for AI training and deployment.</li><li dir="ltr">How Do NPUs and TPUs Compare to GPUs? NPUs are specialized for efficient AI/ML tasks, particularly inference, and are optimized for low-power edge devices like smartphones. TPUs excel in tensor-based operations for deep learning, offering high-speed training and inference, primarily in Google's ecosystem. GPUs, versatile in design, handle a wide range of workloads, including graphics rendering, ML training, and scientific computing, making them ideal for parallel processing. While GPUs are general-purpose, NPUs and TPUs focus on AI-centric optimizations.</li></ol>Recommended reading:&nbsp;<a href="https://www.wevolver.com/article/understanding-nvidia-cuda-cores-a-comprehensive-guide" target="_blank">Understanding Nvidia CUDA Cores: A Comprehensive Guide</a><h2 dir="ltr">References</h2>[1] Google. Cloud TPU documentation. Available from: <a href="https://cloud.google.com/tpu/docs" target="_blank">https://cloud.google.com/tpu/docs</a>[2] Apple. Deploying Transformers on the Apple Neural Engine. Available from: <a href="https://machinelearning.apple.com/research/neural-engine-transformers" target="_blank">https://machinelearning.apple.com/research/neural-engine-transformers</a>[3] Van Belle JL. Why neural processing units (NPUs) are the next Big Thing in AI. 2024. doi:10.13140/RG.2.2.35619.87845. Available from: <a href="https://www.researchgate.net/publication/383184868_Why_neural_processing_units_NPUs_are_the_next_Big_Thing_in_AI" target="_blank">https://www.researchgate.net/publication/383184868_Why_neural_processing_units_NPUs_are_the_next_Big_Thing_in_AI</a>[4] Sanmartín D, Prohaska V. Exploring TPUs for AI Applications. IE Robotics &amp; AI Club, IE University, Madrid, Spain. arXiv preprint arXiv:2309.08918. Available from:<a href="https://arxiv.org/pdf/2309.08918" target="_blank">&nbsp;https://arxiv.org/pdf/2309.08918</a>.

NPUs are integrated units that excel in real-time AI tasks on edge devices like smartphones and IoT systems with low power consumption. TPUs are standalone processors designed for large-scale AI workloads in data centers, delivering exceptional performance in deep learning tasks.

NPU vs TPU: Understanding the Key Differences in AI Hardware Accelerators

Data centers form the backbone of countless digital services, from cloud computing to artificial intelligence (AI). However, as data center infrastructures scale up, they face an increasing number of challenges. First, the rapid adoption of AI is accompanied by significant growth in data volume and increased computing requirements. <a href="https://www.statista.com/statistics/871513/worldwide-data-created/">By 2025, data volume is expected to reach 180 zettabytes</a>, compared to 15 zettabytes in 2015. Additionally, the demand for <a href="https://openai.com/index/ai-and-compute/">computing power has been doubling every 3.4 months since 2012</a> to train cutting-edge AI models. This continuous increase in computational power directly impacts electricity consumption, with AI data centers expected to account for up to 7% of global electricity demand by 2030.Another significant challenge related to data center infrastructure is power failure. Power shortages in data centers can lead to costly downtimes, data loss, and even equipment damage. To prevent power failures, ensuring power reliability is a critical concern. In response to the demand for stable power supplies, Infineon has introduced its&nbsp;<a href="https://www.infineon.com/cms/en/applications/solutions/system-diagnostics-analytics/power-supply-health-monitoring/power-system-reliability-modeling/?redirId=274819">Power System Reliability Modeling (PSRM</a>), a comprehensive solution aimed at enhancing power reliability and reducing the likelihood of outages in data centers and telecom infrastructure by empowering customers accessibility to in-field power supply real-time data.This article will delve into the importance of PSRM and how it could transform data centers companies interpretation of reliability in power, leading to cost savings, improved sustainability and increased resiliency.<h2>The Growing Power Demand and Risk in Data Centers</h2>Data centers house thousands of servers, networking devices, and other critical systems, consuming vast amounts of electricity. As technologies like AI, machine learning, and neural networking advance, the demand for more computational power grows exponentially. The increasing power density in data centers amplifies the risk of power interruptions. Even a brief outage can disrupt services, potentially affecting millions of users and costing companies significant amounts of revenue. Power inefficiencies also increase operational costs and elevate the carbon footprint of these facilities.Given these risks, the ability to maintain uninterrupted power supply is paramount, and this is where Infineon’s new PSRM comes into play.<h2>What is Power System Reliability Modeling?</h2>Infineon’s Power System Reliability Modeling is a solution that combines hardware and software to enhance the reliability and efficiency of power supplies in data centers. It provides a comprehensive view of the entire power system and predicts potential points of failure before they occur. The solution strengths lie in its predictive maintenance enablement capabilities, enabling companies to address in-field issues before they lead to a failure or outage. By utilizing real-time monitoring of the mission profile of the power supply unit, it offers a dynamic understanding of the power supply’s current health status and how it effects its long-term reliability.The PSRM solution focuses on several core insights:<ul><li>Real-time Power Monitoring: mission profile logging features enable real time tracking of operating conditions of the power systems such as input voltage, output voltage, load current, and temperature. By modelling these parameters with component´s FIT rate, we can assess reliability conditions based on the actual mission profile. This allows to anticipate anomalies or fluctuations that could indicate potential problems.</li><li>Lifetime Estimation: By analyzing the mission profile over time, one can predict the operational lifespan of different power components, allowing operators to replace parts before they fail and balance the load in redundant systems, secure highly reliable system operations.</li><li>Predictive Maintenance: The collected modelled data allows data center operators to carry out maintenance tasks and optimize fleet management operations based on actual performance data, rather than fixed schedules. Data-driven decisions can optimize operational efficiency and reduce unnecessary costs.&nbsp;</li></ul><h2>Benefits of Power System Reliability Modeling</h2><h3>Reduced Downtime and Operational Costs</h3>One of the most significant advantages of implementing this solution in IT power distribution systems lies in enabling companies to use data to minimize unexpected downtimes and prevent disruptions in operations. Traditional maintenance models rely on reactive or periodic maintenance schedules, where issues are addressed only after they occur, or parts are replaced based on time intervals rather than actual wear and tear. In contrast, <a href="https://www.wevolver.com/article/predictive-maintenance-using-smart-sensors-to-get-the-most-out-of-your-assets">predictive maintenance</a> models use real-time data and advanced analytics to anticipate failures, allowing for interventions based on the actual condition of components. By leveraging real-time data from the power supply units, the solution helps companies to ensure that potential problems are identified early, thus preventing power interruptions.This predictive approach results in fewer power outages, lower operational costs, and increased system reliability. Data center operators no longer have to deal with the unpredictability of power-related issues, ensuring smoother operations and higher profitability.<h3>Improved Energy Efficiency and Sustainability</h3>Power reliability isn't just about preventing outages; it’s also about optimizing power usage. With the insights provided with the help of Infineon’s solution, data center providers can assess and optimize the energy efficiency of their power systems. This is particularly important in the context of sustainability, where data centers are under increasing pressure to reduce their environmental impact.Hence, data centers can achieve higher energy efficiency by ensuring that their power systems are operating optimally, reducing waste and unnecessary power consumption. In a long run, lower energy use means lower carbon emissions, helping companies meet their sustainability goals while also cutting operational costs.<h3>Enhanced System Reliability&nbsp;</h3>Reliability modeling on the system level empowers customers' power supply operation and real-time monitoring capabilities. The solution will shed light on vital insights into the complete system health status and proactively allow optimization of its usage to extend the operational lifetime. By leveraging these capabilities, customers can manage their power supply systems based on system health, ensuring enhanced reliability and prolonged operational continuity for critical infrastructure. The key benefits of enhanced system reliability provided by the solution can be summarized as:<ul><li>Real-time system diagnostics to any power supply</li><li>Powerful system reliability-centric value proposition</li><li>Seamless integration into power designs through power controllers.</li><li>Easy usage</li></ul><h3>The Future of Data Center Power Management</h3>Infineon’s Power System Reliability Modeling is a crucial innovation at a time when data centers are becoming more integral to global infrastructure. With the rise of AI, cloud services, and the Internet of Things (IoT), the demand for reliable, efficient, and sustainable power systems is greater than ever. Infineon’s solution directly addresses these needs, offering a way to enhance power system resilience while also reducing operational costs and environmental impact.The introduction of PSRM underscores Infineon’s commitment to providing cutting-edge solutions that go beyond simple component manufacturing. By integrating hardware and software, the company offers a holistic approach to power system reliability modelling when designing the next gen of power supplies. Hence, it supports the increasingly complex needs of modern data centers manufacturers and users.<h2>Conclusion</h2>As data centers continue to expand in size and complexity, innovations like <a href="https://www.infineon.com/cms/en/applications/solutions/system-diagnostics-analytics/power-supply-health-monitoring/power-system-reliability-modeling/?redirId=274819">Infineon’s Power System Reliability Modeling</a> will be critical in ensuring that they remain resilient, profitable, and sustainable in the future. The solution enables real-time power supply health monitoring and lifetime estimation, helping to reduce downtime, maximize device usage, lower operational costs, and improve energy efficiency.&nbsp;Finally, it should not be overlooked that Infineon offers a comprehensive array of additional <a href="https://www.infineon.com/cms/en/applications/information-communication-technology/hyperscale-computing/">solutions</a> for AI data centers, covering everything from the grid to the core. The innovative technologies represent a significant step forward in addressing the power challenges, and the offered solutions are poised to redefine the way data centers operate.&nbsp;

Infineon’s Power System Reliability Modeling enhances power reliability in data centers by enabling real-time power supply monitoring, predictive maintenance, and lifetime estimation from component to system level. 

Power System Reliability Modeling: A Step Toward Resilient Data Centers

Random Access Memory (RAM) strip inserted into the Motherboard

This article is a detailed analysis of In-Memory Compute technology, covering its architecture, use cases, recent advancements, and practical implementation strategies to enhance computational efficiency.

In Memory Compute: Transforming Data Processing for Speed and Scalability

Large Language Models such as OpenAI’s ChatGPT have already transformed the way many of us go about some of our daily tasks. These generative artificial intelligence chatbots are trained with language — hundreds of terabytes of text ‘scraped’ from across the Internet and with billions of parameters.Looking ahead, many believe the ‘engines’ that drive generative artificial intelligence will be multimodal models that are not just trained on text but also can process various other modalities of information, including images, video, sound, and modalities from other domains such as biological or atmospheric data.Yet, until recently, training a single model to handle a wide range of modalities – inputs – and tasks – outputs – faced significant challenges. For example, the training often led to a reduction in performance compared to single-task models and typically required careful strategies to reduce quality losses and maximize accuracy. In addition, training one network on different modalities - or inputs - such as language, images or videos that vary greatly, presented additional complexities, and essential information in certain modalities was often incorrectly ignored by the model.<h2>Multimodal Modeling</h2>In a multi-year project undertaken with support from Apple in California, EPFL researchers from the <a href="https://vilab.epfl.ch/">Visual Intelligence and Learning Laboratory</a> (VILAB) in the <a href="https://www.epfl.ch/schools/ic/">School of Computer and Communication Sciences</a> (IC) have developed 4M, for Massively Masked Multimodal Modeling, one of the world’s most advanced single neural networks to handle a wide and varied range of tasks and modalities.In their latest<a href="https://4m.epfl.ch/#fourm21_paper">&nbsp;research paper</a> on 4M, presented in December at NeurIPS 2024, the Annual Conference on Neural Information Processing Systems, the researchers describe how it expands the capabilities of existing models in multiple ways (see box below for more technical details).“With 4M, we now have a rich model that can interpret more than just language. But why does this matter? One common criticism of LLMs is that their knowledge is not grounded because the training data is limited to only language,” explained Assistant Professor Amir Zamir, Head of VILAB.“When we advance to multimodal modeling, we don’t have to limit ourselves to language. We bring in other modalities, including sensors. For example, we can communicate an orange through the word ‘orange,’ just like in language models, but also through a collection of pixels, meaning how the orange looks, or through the sense of touch, capturing how touching an orange feels. If you assemble various modalities, you have a more complete encapsulation of the physical reality that we are trying to model,” he continued.<img src="https://s3-us-west-1.amazonaws.com/wevolver-project-images/froala%2F1736298018585-bc5ffbec.jpg" style="width: 100%px;" class="fr-fic fr-dib">Trying to model the physical reality by assembling various modalities: the image shows a couple of oranges seen through the lens of multiple modalities, with each slice showing a different way one might perceive and understand this scene. The modalities from left to right represent surface normals (the color represents surface orientation), depth (distance to the camera, red=near, blue=far), RGB (the original image), segmentation (distinct objects and image regions), and edges (object or texture boundaries). 2025 EPFL/Visual Intelligence and Learning Laboratory - CC-BY-SA 4.0<h2 id="isPasted">Towards an open-source, generic model for wide use</h2>Despite these impressive advances, Zamir says the development of 4M has presented some intriguing challenges, including the model not developing a truly unified representation across the modalities, and he has his own theory as to why.“We think that secretly, under the hood, the models cheat and create a little ensemble of independent models. One set of parameters solves one problem, another set of parameters solves another, and collectively, they appear to solve the overall problem. But they’re not truly unifying their knowledge in a way that enables a compact joint representation of the environment that would be a good portal to the world.”The VILAB team is continuing to work on building more structure and unification into 4M, with the goal of developing an open-source, generic architecture, enabling experts in other domains to adapt it to their specific needs, such as climate modeling or biomedical research. The team also works on addressing other important aspects, such as boosting the scalability even further and methods for the specialization of models to deployment contexts.“The whole point of open sourcing is that people can tailor the model for themselves with their own data and their own specifications. 4M is coming at the right moment in time, and we are especially enthusiastic about other domains adopting this line of modeling for their specific use cases. We are excited to see where that leads. But there are still a lot of challenges, and there is still a lot to do,” said Oguzhan Fatih Kar and Roman Bachmann, Doctoral Assistants in VILAB and co-authors of the paper.Based on the team’s experience developing 4M and the intriguing problems that they continue to work on, Zamir believes there are some interesting questions around the future development of foundation models.“As humans, we have five key senses, and on top of that, we efficiently learn language, which adds labels and structure to the knowledge that was already grounded in these other senses. It’s the opposite with the current AI – we have language models without sensory access to the world but that are trained using colossal data and compute resources. Our goal is to study the role of multimodality and efficiently develop a grounded world model that can be effectively utilized for downstream uses.”<hr>4M expands the capabilities of existing models across several key axes, including:<ul><li>Modalities: 4M enables new capabilities like predicting tens of modalities from tens of others, cross-modal retrieval, controllable generation, and strong out-of-the-box performance. It has convincingly shown that a single model can solve tens of diverse tasks without any loss in performance compared to dedicated single-task models and the state-of-the-art.</li><li>Diversity: 4M supports varied modalities and more structured data, such as human poses, SAM instances, and metadata for controllable generation.</li><li>Tokenization: 4M investigates discrete tokenization of diverse modalities such as global image embeddings, human poses, and semantics.</li><li>Scale: The public model has been scaled to 3 billion parameters and trained on over 500 billion tokens.</li><li>Co-Training: 4M demonstrates co-training on vision and language modeling simultaneously.</li></ul>

A couple of oranges seen through the lens of multiple modalities, with each slice showing a different way one might perceive and understand this scene - 2025 EPFL - CC-BY-SA 4.0

EPFL researchers have developed 4M, a next-generation, open-sourced framework for training versatile and scalable multimodal foundation models that go beyond language.

big data

Latest Posts