Polyn has developed an Analog Neural Network Chip
21 June, 2022
The new concept is based on a mathematical discovery that allows for the representation of digital neural networks using operational amplifiers and resistors - to achieve high speed and lower power consumption
Polyn Technology plans to introduce a novel Neuromorphic processor chip, based on analog electrical circuitry, unlike the standard digital neural networks. Lately, GlobalFoundries has completed the production of the first trial batch in a 55nm CMOS process. Alexander Timofeev, CEO and co-founder of Polyn Technology told Techtime that the chip’s commercial introduction is planned for November or December 2022. The company’s NASP (Neuromorphic Analog Signal Processing) technology had started as a mathematical development of the Chief Scientist and co-founder Dmitry Godovsky. He developed an equation that makes it possible to represent inferencing neural networks using analog elements such as operation amplifiers (Opamps) and resistors.
Based on that concept, Polyn was founded in 2019, and recruited Israeli developers from IBM, Intel, Tower Semiconductors, Cadance etc. The company successfully implemented the new representation by building Analog Neuromorphic Model composed of thousands of digital neurons using classic building blocks: Operational Amplifiers (OpAmp) and Resistors. Polyn’s first NASP chip contains 50,000 analog neurons. Timofeev estimates that its power consumption is 100 times better compared to a parallel digital neural network, and 1,000 times faster.
A combination of Digital Network and an Analog Processor
Since its foundation, Polyn raised $4.5 million, and currently it employs 22 employees. Although registered in UK and holds an office in London, Polyn’s headquarters is located in Caesarea, Israel. It is currently in the process of registering 21 different patents to protect its technology. The final product is a hybrid of a fixed analog network for pattern identification, and additional dynamic digital component in charge of interpreting patterns.
This component is also responsible for the addition of new components to the algorithm during updates and even to perform training at some level. Timofeev: “An average of 90% of the neural network is fixed and unchanged. Less the 10% is updated during operation. Our chip is hybrid: we transform 90% of the network to an analog circuit, and the rest remains digital. We can also adjust this ratio according to the specific needs.”
Three Chips Roamap
The first processor, NeuroSense, will be the first member of NASP family and is planned to reach the market by the end of 2022. Its power consumption is below 100µW, which make it perfect for low-consumption accessories such as smart watches and wearable devices. It will be marketed as a physical component and as an Intellectual Property (IP) offer. “NeuroSense will solve an existing problem within all smart watches: They have to constantly monitor all sensors, something that consume lots of energy.”
Polyn’s roadmap contains two more components: An Audio Processor, intended to segregate human voice from background voices, and is designed for the hearing aids market, first responders teams, radio communication devices and later also consumer applications. The third chip will focus on shock analysis for the industrial market.
What is the immunity level of the chips?
“We rely on existing, well proven production processes. This is why the production of the first chip is performed in CMOS 55nm technology. We perform 50 inferences per second. This is very low frequency, which makes the component highly immune to electromagnetic interferences. As any other neural network, it is a parallel network, meaning a fault in single neuron does not affect the final result. This is not a circuit that accumulates faults.”