SiFive and BrainChip have partnered to show their IP is compatible in SoC designs for embedded artificial intelligence (AI). The companies have demonstrated BrainChip’s neuromorphic processing unit (NPU) IP working alongside SiFive’s RISC–V host processor IP.
Brainchip’s NPU processor IP, the basis for its Akida chip, is a neuromorphic processor designed to accelerate spiking neural networks. This IP can be used to analyze inputs from most sensor types, including cameras, to provide ultra–low power analysis in real–time applications. A recent BrainChip demo showed its Akida chip in a vehicle, detecting the driver, recognizing the driver’s face, and identifying their voice simultaneously. Keyword spotting required 600 µW, facial recognition needed 22 mW, and the visual wake–word inference used to detect the driver was 6–8 mW.
SiFive is a provider of RISC–V processor IP, including its Intelligence series of multi–core capable RISC–V processors with vector extensions which are optimized for AI workloads in edge devices.
“BrainChip can run [AI] algorithms on their own, but when they move into a larger system, they will need a host processor,” Chris Jones, vice president, product at SiFive, told EE Times. “You could pick a host processor that does nothing but scheduling, or you could pick a host processor that actually contributes to the AI processing, and that’s where the SiFive Intelligence product comes in.”
In an SoC design for edge AI, the AI workload would typically be split between host processor, vector processor, and AI accelerator — some parts of edge workloads are better suited to general purpose compute rather than a dedicated AI accelerator, Jones said.
“It’s advantageous for BrainChip to align with industry leaders to make sure their customers have a seamless integration experience, so BrainChip can deliver the requisite software that runs on the host processor and makes it easier for the end user to integrate their products and ours,” he said.
Jones described work done so far as the “tip of the iceberg,” adding that the two companies have so far demonstrated compatibility of BrainChip’s IP with SiFive’s RISC–V architecture. The companies will work together on an ongoing basis to further integrate software and hardware IP.
“We have ambitious plans going forward,” Jones said. “SiFive has made great strides in the last year or so, bringing vector processing to market now that [vector processing] has found its niche in AI and image signal processing.”
Part of SiFive’s plan is to build an ecosystem of AI accelerator IP providers whose products are compatible with its host processor IP.
“Our relationship with BrainChip is in no way exclusive,” Jones said. “BrainChip is the first partner we’ve gone public with, but we’re talking to many other players in this space.”
While BrainChip is the company’s first hardware IP partner, SiFive has design wins for its host processor IP, including data center AI accelerator company Tenstorrent.
“Certainly, we’re open to exploring partnerships with people who have novel technology,” he said. “Companies that are bringing IP or chips to market can’t ignore RISC–V as a platform just because of the market share we’re taking.”