For a while now, NVIDIA has been working to turn its NVLink high speed interconnect technology into more than just a bus used by NVIDIA, aiming to fashion an entire ecosystem around it. Those efforts kicked into high gear last year with the announcement of NVLInk Fusion, a program to allow third parties to connect to other NVLInk hardware and even full NVLink networks via a combination of IP licensing and external chiplets, giving those vendors all the tools needed to add NVLink to their chip designs.
Since then, NVIDIA has slowly been winning over chip and IP vendors. Following Arm, Intel, and AWS jumping on board last year, this morning the NVLink Fusion ecosystem is gaining its first RISC-V vendor: SiFive. The high-performance CPU IP firm has announced that it is joining the NVLink Fusion ecosystem by adopting NVLink Fusion for its future data center-class chips. As a result, going forward it is going to become possible to plug in Arm, x86, and now RISC-V CPUs in to NVIDIA’s data center GPUs via the NVLink interface, giving system builders a wider suite of options for CPUs to use in AI systems.
Digging a bit deeper, the full announcement from SiFive (and with supporting quotes from NVIDIA) is relatively brief and does not announce any specific product plans beyond adding NVLink support to SiFive’s “high-performance data center-class solutions.” Presumably we’re looking at hardware in the Vera Rubin platform timeframe, if not beyond, which means they’re implementing NVLInk 6 generation technology.
Reading between the lines, as a CPU IP vendor, SiFive’s interest in NVLink would be in using the C2C variant of NVLink technology for their chips. The chip-to-chip version of NVLink offers a high-bandwidth fully cache-coherent link between the CPU and GPU, and it is the preferred way to connect to NVIDIA’s GPUs in highly-integrated systems – already being the technology NVIDIA uses to wire up its own Grace (and Vera) CPUs to their GPUs. As part of the NVLink Fusion ecosystem, NVIDIA offers NVLInk-C2C on an IP licensing basis, so SiFive will be able to integrate the necessary IP straight into their future chip designs.
For SiFive, adopting NVLink Fusion support will open the door to RISC-V CPUs based on the company’s IP being used in high performance AI systems. The company has long argued it has an edge in this space due to the efficiency and customizability of its designs – but with NVIDIA owning the lion’s share of the AI accelerator market, they were always going to be on the outside looking in without the ability to interface with NVIDIA’s GPUs via the higher performing NVLink-C2C interface. As well, this gives them a leg up in a space with plenty of other competitors, especially given RISC-V’s open ISA.
As for NVIDIA, this is a win by not only further expanding the NVLInk Fusion ecosystem, but by securing a major RISC-V firm the company can hedge its bets with regards to CPU support. If customers don’t want to build integrated AI systems with x86 or ARM CPUs, this will give them the option to use RISC-V instead of stepping back from the NVIDIA ecosystem entirely. This agreement follows on the heels of NVIDIA’s RISC-V announcement last year, where the company announced that they’d be bringing CUDA and NVIDIA’s drivers to RISC-V.
