Welcome

Low-cost, low-power FPGA-based MPUs that supercharge your AI development

Find out how we are pushing the boundaries of what's possible

Explore Now

Welcome

Low-cost, low-power FPGA-based MPUs that supercharge your AI development

Find out how we are pushing the boundaries of what's possible

Explore Now

About Nsquare

AI & Sensors

Good quality data is fundamental to developing effective, reliable, and fair AI systems. It is of even greater importance when developing next-gen AI technologies that will autonomously drive cars, transport freight across the oceans, guide robotic manufacturing in factories, or proactively run the data centers of tomorrow. For such advanced systems to be truly brought to life in a safe, predictable, and profitable manner, the physical world must first be studied and quantified in near-perfect detail.

Studying the physical world is a task that has long been performed by a class of devices called sensors. These detect and measure physical properties, such as temperature, pressure, light, or motion, and convert these measurements into electrical signals (i.e., data) that can be interpreted, recorded, or acted upon by other devices. Over the past century, sensor technology has made some of the technological staples of the modern world possible – such as the humble automatic door, the ubiquitous smoke alarm, or the advanced automotive airbag.

New Challenges

While sensors keep doing what they have always been doing - converting physical information into data - the reasons why we collect data have themselves evolved significantly. The traditional use of sensors is adequate when building a vending machine, sprinkler system, or even an assembly line. However, when it comes to training the next generation of AI models, these workhorses of science & industry need to evolve to entirely new levels of sophistication.

In the real world, the presence of random noise in the environment (called Gaussian) is a guarantee. This is especially problematic for AI development because random noise (i.e., random data) by its very defintion is statistically uncorrelated. Since modern AI models (machine learning) are at their core statistical pattern-recognition systems, random noise in the training dataset directly degrades their overall prediction accuracy.

Presently, separating noise from actual, meaningful data can consume a large volume of data center processing. For training the most effective models, ensuring the quality of the dataset is a must.

High-Quality Data For High-Quality AI

High-quality, noise-free data provides better insights and analyses into any available dataset. A better understanding of the data then helps train AI models that are more informed, allowing for more accurate conclusions to be drawn in real time. For a car driving autonomously, this can mean the difference between recognizing a sound as simply traffic noise, or the sound of an imminently approaching emergency vehicle.

In contrast, poor-quality data can introduce uncertainty and biases in AI models. In addition to hurting the model effectiveness and its technological outcomes, it can undermine the trustworthiness of AI-driven decisions as a whole.

Low-Cost Solutions For High-Quality Data

While the giants of industry (NVIDIA, AMD, etc.) focus on making the ultimate AI accelerators, there is significant unmet need for AI solutions in the less glamorous but still critical tasks of real-world sensor data acquisition. These are performed daily at virtually all critical facilities in the world such as factories, power plants, airports, and even places like supermarkets and fast-food restaurants.

Specifically, we have identified a market opportunity for a low-cost, low-power chip using advanced mathematical techniques to reject statistical noise even before it gets recorded. Acting as a middle layer between traditional sensors and the connected computers, MPUs (Mathematical Processing Units) will analyze the incoming information using a host of advanced mathematical techniques such as Kalman Filtering, Bayesian Inference, or Fourier Transforms, and transmit only the relevant/processed information along to the reading device.

Benefits & Opportunities

Using an MPU in the data acquisition stages will free up precious computational resources in already overextended data centers, and ease the exponentially growing demand for electrical power to run these systems. If implemented correctly, MPUs present a significant opportunity to save on both processing costs and power consumption.

There is additional opportunity as the devices making up the Internet-of-Things (IoT) are projected to explode in number and become accepted and expected parts of the average person’s life. Instead of the present approach of utilizing data centers to gather and then process large amounts of sensor data, why not have the data be already-useful before it reaches your servers?

When combined with traditional sensors and existing AI accelerators, MPUs promise a future where AI models can be trained, tested, and used, all while remaining completely untethered from the internet and its vast computational power, or even from the nearest possible power outlet.

Our Goal

Our goal is to implement an MPU (or family of MPUs) using Field-Programmable Gate Arrays (FPGAs). These are versatile, low-cost semiconductor devices that can be customized to perform various digital functions. Unlike fully custom chips, FPGAs are manufactured in large volumes and can be configured to match specific use cases.

The immediate goal is to implement an MPU prototype that performs real-time Kalman Filtering on real-world data. Kalman Filtering is one of the many mathematical techniques used extensively in the industrial and scientific fields to estimate the state of a dynamic system from a series of noisy measurements.

Prototype Demonstration

We then test this MPU with a portable hand-held testbed, called a Tricorder. This custom-made device is a 'pod' of 23 commonly used commercially-available sensors in a portable, hand-held form factor. It reads a whole host of data and transforms it into powerful visualizations that help its user make better sense of the world.

Witness in real-time as our MPU drastically improves the quality of its data acquisition, and how this makes an effect on the end use case, which in this case is design, exploration and learning, but more often is critical industrial, scientific, or military applications.

Learn More

Meet The MPU: Mathematical Processing Unit

The next generation of high-quality real-time data acquisition

Implemented using commodity FPGAs, MPUs are capable of implementing several advanced mathematical techniques in real time. They will act as a middle layer between commodity sensors and recording devices with the exact type of computation and its parameters being user selectable, depending upon the use case. This allows flexibility to work with multiple sensor and data types.

Read about some of the advanced mathematical techniques that can be implemented in real time by MPUs:

Fourier Transform

Use Case: Signal processing, speech recognition, and image processing.

Example: Implementing a Fast Fourier Transform (FFT) on an FPGA can accelerate the processing of signals in real-time applications. In AI workloads, FFTs are often used for frequency analysis, filtering, and feature extraction from time-series data. For instance, in speech recognition, FFTs can be used to convert audio signals into the frequency domain for more efficient processing.

Convolution

Use Case: Image and video processing, object detection, and computer vision.

Example: Convolution operations are fundamental in CNNs (Convolutional Neural Networks). By implementing convolution operations directly on an FPGA, it’s possible to accelerate the processing of large-scale image data for tasks like object detection, segmentation, and classification. FPGAs can be configured to perform convolutions in parallel, reducing the time required to process each layer of a CNN.

Matrix Multiplication

Use Case: Neural networks, deep learning, and machine learning algorithms.

Example: Matrix multiplication is at the core of many machine learning algorithms, especially in deep learning for operations such as fully connected layers, recurrent neural networks (RNNs), and transformers. FPGAs can implement highly parallelized matrix multiplication units, making them ideal for accelerating deep learning inference tasks.

Gradient Descent

Use Case: Training machine learning models, optimization algorithms.

Example: Gradient descent is used to minimize the loss function during the training of machine learning models. Implementing gradient descent algorithms on an FPGA can accelerate the training process, particularly in scenarios where custom optimization methods are used. This is beneficial in real-time or low-power AI training applications.

Sparse Matrix Operations

Use Case: Natural language processing (NLP), recommendation systems, and graph-based algorithms.

Example: Sparse matrix operations are common in NLP for handling large, sparse datasets (e.g., word embeddings, recommendation matrices). FPGAs can be optimized to handle sparse matrices efficiently, reducing memory usage and computation time, which is critical for scaling AI workloads in these areas.

Bayesian Inference

Use Case: Probabilistic reasoning, decision making, and robotics.

Example: Bayesian networks and other probabilistic models can be implemented on FPGAs to accelerate decision-making processes in AI systems. This is particularly useful in robotics, where real-time decision-making is required, and in AI systems that need to reason under uncertainty.

Support Vector Machines (SVM)

Use Case: Classification, regression, and anomaly detection.

Example: FPGAs can be used to implement the core operations of SVMs, such as dot products and decision functions, to accelerate classification tasks. SVMs are commonly used in scenarios where high accuracy and real-time performance are required, such as in fault detection and biometric recognition.

Principle Component Analysis (PCA)

Use Case: Dimensionality reduction, feature extraction, and data compression.

Example: PCA is used to reduce the dimensionality of data while retaining most of the variability in the dataset. Implementing PCA on an FPGA can accelerate the preprocessing stage in AI pipelines, particularly in applications involving large datasets, like image recognition or genomics.

Markov Decision Process (MDP)

Use Case: Reinforcement learning, decision making, and autonomous systems.

Example: MDPs are used in reinforcement learning to model decision-making in environments with stochastic outcomes. FPGAs can implement MDPs to accelerate the computation of value functions and policy iteration, which is crucial in applications like autonomous vehicles and robotics.

Numerical Solvers (ODEs/PDEs)

Use Case: Physics-based simulations, weather modeling, and financial modeling.

Example: Numerical solvers for ordinary differential equations (ODEs) and partial differential equations (PDEs) are used in simulations and modeling tasks. Implementing these solvers on FPGAs can accelerate simulations in AI systems that require real-time analysis, such as in financial markets or climate prediction models.

Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) Networks

Use Case: Sequence prediction, time-series analysis, and NLP.

Example: FPGAs can implement the key operations of RNNs and LSTMs, such as matrix multiplications and activation functions, to accelerate the processing of sequential data. This is particularly useful in NLP tasks like language modeling and machine translation.

Low-cost, low-power FPGA-based MPUs that supercharge your AI development

Low-cost, low-power FPGA-based MPUs that supercharge your AI development

About Nsquare

AI & Sensors

New Challenges

High-Quality Data For High-Quality AI

Low-Cost Solutions For High-Quality Data

Benefits & Opportunities

Our Goal

Prototype Demonstration

Meet The MPU: Mathematical Processing Unit

The next generation of high-quality real-time data acquisition

Fourier Transform

Convolution

Matrix Multiplication

Gradient Descent

Sparse Matrix Operations

Bayesian Inference

Support Vector Machines (SVM)

Principle Component Analysis (PCA)

Markov Decision Process (MDP)

Numerical Solvers (ODEs/PDEs)

Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) Networks

This website uses cookies.