Denso partners with Quadric to develop NPU

Japanese automotive component supplier Denso, has signed a development license agreement with Californian startup, Quadric for a neural processing unit (NPU).

The agreement means Denso will acquire the IP core license for Quadric’s Chimera general purpose NPU (GPNPU). Both companies will co-develop IP for an in-vehicle semiconductor for use in intelligent vehicle systems (ADAS and AD/self-driving) and inter-vehicle and cloud communications.

Specifically, Denso will combine the Quadric GPNPU IP along with a Denso-designed RISC-V supervisory processor into an integrated subsystem that meets ASIL-B/D safety standards. Both companies are collaborating on the integration, testing, software development and safety qualification process.
The Chimera GPNPU enables system developers to add their own AI capabilities and updates in software. It is also claimed to have “dramatically lower power than alternative GPU designs”.

“In recent years as companies have raced to develop ADAS/AD algorithms, a common approach for development and testing has been for automotive OEMs and Tier 1 suppliers to put a rack of high-powered data centre GPUs in a modified vehicle trunk to quickly get test systems on the road,” explained Steve Roddy, Quadric’s chief marketing officer. “But having 800W, $5000 systems in the vehicle is impractical in volume production. Data centre GPUs employ architectures that rely upon cache-based data memory systems, graphics DDR memory or HBM memory, full floating point (FP32) data formats, and assumptions of boundless data storage. Embedded SoCs by contrast need to run at much lower power: 5W to 10W range – while still delivering 100s of TOPs of ML inference performance.”
Denso, a Tier 1 automotive supplier subjected the silicon-proven GPNPU to intense benchmarking “for several years”, with the intention of implementing the cores into a Denso-commissioned SoC design. “We have high expectations for the development of in-vehicle semiconductor IP that can flexibly respond to the AI trend with low power by utilising Quadric Chimera GPNPU,” commented Hiroshi Kondo, head of Mobility Electronics Group, Denso.

The development cycle for automotive applications will be several more years, said the company, although Quadric customers with non-automotive, consumer-grade products are likely to come to market sooner.