Neuromorphic Front-End Chip
NASP technology for near-sensor data pre-processing
Many applications could benefit from the neural network paradigm, but practical implementation of this formidable mathematical method suffers from excess power consumption when performed in the traditional way on standard CPUs or GPUs. If an application uses a large amount of data, and accesses the memory very often, it causes a bottleneck within Von Neuman architecture. For cases with a continuous signal flow, special purpose processors are more efficient.
For devices that perform truly always-on measurements, Neuromorphic Analog Signal Processing (NASP) is an ideal solution with ultra-low 100uW power consumption.
If all sensor data must be sent to a centre for analysis, the data communication and processing would be more trouble than it’s worth. On-sensor data pre-processing could significantly reduce the amount of data sent to the cloud, saving money, improving latency and user privacy.
The NASP addresses all such situations with smart optimization (pre-processing) of raw data on-device. Not only could it solve problems for existing applications, but also open new opportunities for the whole industry.
On-sensor data optimization
NASP is a true Tiny AI solution targeted for raw data optimization and amount of data forwarded to the cloud. The NASP chip is located right next to a sensor, forming the Tiny AI logical layer. It is an inference solution that uses already trained machine-learning models to make predictions.
In the NASP concept, the data processing chip is synthesized from already trained neural networks by NASP automation tools. Based on POLYN’s years of expertise, this approach is highly efficient for applications such as voice extraction, sound/vibration processing, measurements on wearables and more. It provides a huge advantage in power, accuracy, and latency.
Neuroscience behind the NASP
The main advantage of neuromorphic computing is parallel operation achieved through hardware and software design that strives to mimic the human nervous system and have the highest compute-to-power-consumption efficacy. Besides low power consumption and improved performance of computing workloads, neural networks provide fault tolerance, which means the system can still produce results if a sensor data is inconsistent.
All sensor signals entering the input layer of the NASP chip at the same time are transmitted to the successive layers in parallel. There are no execution cycles, and no instructions directed to/from memory.
The human brain is not only an ultra-low-power parallel operating system, but also an analog system, processing a variety of signals without converting them into a binary format. For tasks such as signal perception, analog systems are preferable. According to Semiconductor Research Corporation, the analog signal deluge is expected in the coming decade, demanding fundamental breakthroughs in hardware to generate smarter world-machine interfaces.
The NASP is precisely one of these breakthroughs, built to perceive analog signals as well as digital ones and, most importantly, to add “intelligence” to various sensors.
The NASP application-specific chips contain artificial neurons (nodes performing computations) and axons (connections with weights between the nodes) implemented using circuitry elements: neurons are implemented using operational amplifiers, and axons by using thin-film resistors.
The digital transformation that the industry is going to embrace will not be possible without an analog computing renaissance for several reasons. One is the concept of energy saving. Excessive power consumption is incompatible with data computations in sensory systems. The next trend is that AI is moving more and more towards the edge and being applied today to sensor nodes. It is required to optimize communications between billions of IoT devices, and to offload data processing from the cloud, improving TCO and efficiency.
Like the human brain, which excels at processing information that is complex, and changing dynamically in time, the Neuromorphic Analog Signal Processors excel in real-time computing, thus contributing to the beneficial meshing of digital and analog tech worlds.
NASP chip design automation tools
The NASP chip design embodies the approach of a sparse neural network, with only the necessary connections between neurons required for inference, which means the solution reduces the neural connections significantly and efficiently. In contrast to in-memory designs, where each neuron is connected to each neighbouring neuron, the NASP approach simplifies the chip layout.
A neural network adjustment to the chip design is a significant part of every neural network-on-chip solution. Programmable solutions available today in the market have architectural restrictions that impose additional transformation on a neural network. Sometimes, the original neural network undergoes an almost 100% transformation during porting, which is a costly approach.
To address this issue, the NASP model includes the chip design automation tools, namely POLYN’s T- Compiler and Synthesis tools, that convert any trained neural network into an optimal math model for further chip layout generation, while completely preserving compliance with POLYN’s customer neural network, and saving related efforts and costs.
POLYN Technology is a fabless semiconductor company supplying Neuromorphic Analog Signal Processing (NASP) technology and Tiny AI chips.
NASP solutions are best suited for always-on smart devices performing sensor signal pre-processing in a wide variety of Edge AI applications such as hearables, wearables, smart home, and Industry 4.0. POLYN offers best-in-class performance for any edge IoT device that requires ultra-low power consumption, high accuracy, low latency, and small size.
POLYN brings machine learning to the edge through patented methods of data processing, mathematical simulation, and chip design, all resulting in an unmatched ratio of computing cost to consumed power. At the same time, NASP chips are based on standard CMOS technology and can be both highly tailored to specific applications and yet affordable for the mass market. Thanks to the NASP, many AI/ML workloads that typically would run on high-powered application processors or MEC now can run in NASP-based products to resolve power constraints and benefit from deep learning computations next to the sensors.
The unique NASP technology has a brain-mimicking architecture to handling raw sensor data, including pure analog signals. Moreover, POLYN offers a novel approach to analog integrated circuits aiming to mimic the human nervous system processing in a truly neuromorphic solution.
The NASP uses trained neural networks from any major ML library such as Tensor Flow, PyTorch, MXNet, and others. POLYN assists customers with neural network selection and training. The NASP development framework provides fast and easy conversion of trained neural networks into neuromorphic silicon chipsets. POLYN’s Neural-Net-to-Chip automation tools provide a fully functional math model of a neural network to convert the trained neural network to the chip production files. These unique tools dramatically reduce product time to market, CAPEX, and redesign OPEX.
POLYN offers customers a unique business model that provides customer support throughout the entire product development cycle, including neural network selection and training, optimization and generation of the SW simulation (D-MVP) for the resulting neural network so that the size and structure always best suits the customer’s task. This is POLYN’s fundamental advantage: ensuring optimal solutions and further investment savings for customers. POLYN’s approach supports fast and cost-effective development of tailored solutions that perform deep learning computations on mass-market devices.
NeuroVoice is NASP Tiny AI Voice Extraction chip, addresses the most annoying problems of mass-market hearables: hearing experience in case of irregular noises, power consumption, cloud connectivity and data privacy issues. It is fitted with a neural network enabling clear communication for both a speaker and listener, by app or by phone. The solution features a direct analog interface, no need for complicated integration, and ultra-low power consumption below 100µW.
NeuroSense is POLYN’s Tiny AI chip for wearables, with a fusion of PPG and IMU sensors for the high accuracy heart rate measurement as well as auto-recognition and tracking of human activity. It addresses such well-known problems as short battery life and low tracking accuracy, providing ultra-low power consumption below 100µW and three times better heart rate accuracy than algorithm-based calculations for mass-market fitness trackers, smartwatches, health monitoring devices, and remote care wearables for seniors.
Additional product in the pipeline (VibroSense) targets vibration sensors for predicted maintenance in Industry 4.0 market.
POLYN is led by an international team of professionals experienced in implementing and commercializing new technologies and leading experts in chip design and deep learning.
POLYN Technology was founded in 2019. The company is registered in London and headquartered in Israel.