June 19, 2022
Based on a mathematical discovery that makes it possible to represent digital neural networks using server amplifiers and resistors, and to obtain high speed and power savings. Globalfoundries produced the first chip. Will hit the market at the end of 2022
Company Polyn Technology Caesarea plans to launch by the end of the year a completely new type of neural network processing chip, which is based on an analog electrical circuit and not on a digital neural network. Globalfoundries recently completed the production of the first experimental series, using CMOS technology at a junction width of 55 nanometers. Co-founder and CEO of the company, Alexander Timofeev, told Techtime that the commercial launch of the component is scheduled to take place in November or December 2022.
The company’s Neuromorphic Analog Signal Processing – NASP technology grew out of mathematical development by Chief Scientist Alexander Godowski. He developed an equation that makes it possible to differently represent the activity of neural networks in the heating phase (trained networks). Following the idea, the company was established in 2019, and recruited Israeli developers from companies such as Taur, IBM, Intel, Cadence and more. The company was able to realize the new representation through the construction of analog neurons. This makes it possible to create a circuit that realizes a network of thousands of digital neurons – using classic building blocks: server amplifiers (OpAmp) and resistors.
A hybrid component that combines a digital network and an analog processor
The company’s first NASP chip, for illustration, contains 50,000 analog neurons. According to Timofeev, it is 100 times more cost-effective than a similar digital neural network, and 1,000 times faster. Since its inception, Poland has raised about $ 4.5 million. The company currently employs 22 people. It is registered in the United Kingdom and has offices in London, but is headquartered in Israel. It is currently in the process of registering 21 different patents to protect the technology.
According to Timofeev (pictured below), the final product is a hybrid chip that includes a fixed analog network that is responsible for implementing the pattern recognition phase, and another dynamic digital component such as a small digital neural network, which is responsible for interpreting the patterns, adding components to the algorithm For limited training. Timofeev: “On average, about 90% of the neural network is fixed and does not change. Less than 10% of it receives various updates later in the operation. Our chip is a hybrid: we convert 90% of the network to an analog circuit and the rest remains digital. We can also adapt this mix to the specific application needs. ”
The first processor the company will release at the end of the year, NeuroSense, is planned to be the first component in a family of components. Its power consumption is less than 100 micro-watts, so it is designed for low-power accessories such as smart watches and wearable accessories. It will be sold both in the form of a physical component and in the form of intellectual property (IP). “The chip will solve a problem that exists today in all smartwatches: they need to perform continuous monitoring of the sensors, but this requires a lot of energy.” The company’s roadmap includes two other components: a voice processor to separate the human voice from background noises, and is intended for the market of hearing aids, wireless radios of emergency and rescue forces, and later also for consumer applications. The third chip will focus on shock analysis for the industrial market.
What is the level of immunity of the chips?
“We rely on existing and proven production processes. Therefore, the production of the first chip was also performed at 55 nm. We make 50 inferences per second. This is a very low operating frequency and therefore the component is very resistant to electromagnetic interference. Like any neural network, this is a parallel network where an error in one neuron does not generally affect the end result. This is not a circle that accumulates mistakes. “
Posted in categories: Artificial Intelligence, News, Semiconductors, Israeli Industry
Posted in tags: Polyn, Analog, Artificial Intelligence, Chips
