Featuring an analogue neuromorphic core of a Voice Activity Detection neural network model, the chip will be used to qualify the NASP platform technology and to develop first customer products.
The company’s NASP technology is intended for Edge AI and sensor signal AI processing, and enables cloudless, on-chip AI for various IoT devices while ensuring both privacy and sustainability.
POLYN can facilitate the seamless deployment of edge AI computing across real-time IoT applications, including automotive, consumer electronics, industrial IoT, machine-to-human interfaces, and wearables, among others.
The company’s technology uses analogue neurons with only the necessary connections between neurons for inference. This approach achieves both ultra-low power consumption and latency.
“This is a considerable milestone in NASP chips development toward our mission to bring energy-efficient neuromorphic solutions to market,” said Aleksandr Timofeev, CEO of POLYN. “Our NASP technology platform is one-of-a-kind, and its implementation called for significant engineering ingenuity protected by numerous patents.”
POLYN has designed and implemented a neural network compiler and a tailored EDA flow using standard Cadence tools to transform a trained digital neural network model into mathematically equal analogue neuromorphic cores. This flow efficiently generates a cohesive neural network layout consisting of thousands of neurons and connections with the resistive weights. The flow can be employed at any semiconductor foundry.
POLYN has integrated Cadence Virtuoso for custom analogue design and the Innovus Implementation System for digital place and route, clock tree synthesis, and timing closure with the POLYN proprietary NASP platform – this allowed its engineering team to streamline the chip development process and meet aggressive time-to-market goals.
According to Timofeev POLYN is actively engaged with customers seeking an ultra-low-power solution for voice processing applications.
Additionally, he said that there is high interest in sensor data pre-processing implemented in silicon for automotive industry applications.
