Imagine if a computer's brain could think and remember in the same spot instead of constantly moving data back and forth, which wastes energy. This project builds a chip that does exactly that using a special type of memory that acts like a neural network. It's like replacing a slow courier service between a warehouse and a desk with a desk that has everything built right into the surface.
Edge AI devices are currently limited by a trade-off between computing power and energy consumption. Existing chips are either too power-hungry or lack the maturity for mass production at low cost.
A multiprocessor System on Chip (SoC) prototype in 28nm FD-SOI technology featuring an analog in-memory neural processing unit and RISC-V microprocessor.
If you are a wearable device manufacturer dealing with short battery life due to AI processing — this project developed a 28nm FD-SOI chip that integrates memory and processing. This reduces the energy needed for AI tasks, allowing devices to run longer on a single charge.
If you are an ADAS sensor provider dealing with the need for instant, safe decision-making at the edge — this project developed a functional safe and secure SoC. It provides low-latency AI processing directly on the chip to improve reaction times for safety systems.
If you are a smart factory equipment maker dealing with high costs of deploying powerful AI in remote sensors — this project developed a mass-production compatible PCM technology. This allows for high-performance AI in low-cost, energy-constrained hardware.
Based on available project data, specific pricing is not listed, but the project focuses on using a mature PCM technology to ensure a path compatible with mass volume production and cost.
Yes, the project leverages STMicroelectronics' high-density embedded PCM cell process, which is described as a qualified and mature technology for embedded use in the industry.
Based on available project data, the project involves the design of a modular IP implementing a Neural Processing Unit (NPU), though specific licensing terms are not provided.
The system integrates an analog IMC-based unit with a digital processing subsystem and host subsystems based on an enhanced RISC-V microprocessor implementation.
The project period runs from 2022-09-01 to 2027-02-28, suggesting the pre-product demonstration will be finalized by early 2027.
The consortium is highly balanced for commercialization, consisting of 15 partners with a 47% industry ratio (7 industrial partners, including coordinator STMicroelectronics). The presence of 7 universities and 1 research center ensures a strong R&D pipeline, while the heavy industrial weight suggests a clear focus on translating the 28nm PCM technology into a mass-producible product.
Contact STMicroelectronics SRL regarding the NeuroSoC PCM-based NPU IP
Contact us to explore licensing opportunities for the NeuroSoC IP
