SiFive Aims at RISC-V Generative AI, ML with Its New Performance P870, Intelligence X390 Cores

SiFive, a pioneer in commercialized implementations of the free and open source RISC-V instruction set architecture (ISA), has announced the launch of two new core intellectual property (IP) families — taking aim at powering the future of generative artificial intelligence (AI) and machine learning (ML) workloads.

Announced at an event in Santa Clara, the new SiFive Performance P870 and Intelligence X390 processor core IP are designed to be usable separately — but balance each other out when paired. The Performance P870 is a general-purpose high-performance core, offering a claimed 50 per cent improvement in peak single-threaded performance over the company's prior generation technology — and which comes with the promise of compatibility with Google's official requirements for RISC-V cores to be used with its Android operating system and scalability up to a 32-core cluster.

The Intelligence X390, meanwhile, is the successor to the earlier X380 core which was itself an improvement over the previous-generation X280 — and is designed exclusively for accelerating vector workloads, as found in artificial intelligence and machine learning systems. According to SiFive's in-house testing it delivers a fourfold boost to both performance and bandwidth over the X280 and can be further expanded using the company's Vector Coprocessor Interface Extension (VCIX), which allows customers to add their own instructions or hardware to the core.

Either of the two core IPs can be licensed and used independently, but SiFive is hoping its customers will pick both to use side-by-side for powering high-performance yet energy-efficient generative AI systems. "Together," the company claims, "the new products create a powerful mix of scalar and vector computing to meet the needs of today’s dataflow and computation intensive AI applications across consumer, automotive, and infrastructure markets."

More information on the SiFive Performance and Intelligence processor IP is available on the company's website; at the time of writing, it had not confirmed when the first hardware implementations will land on the market.