Nvidia shuts down software support for RISC-V GPUs

Although Nvidia was an early adopter of open source technology in its GPU controllers, it has not been interested in providing software support for RISC-V-based GPUs.

Nvidia has no plans to add RISC-V support to CUDA, its proprietary GPU software platform, an Nvidia representative told HPCwire in response to questions during the CUDA 12 session at the GTC summit.

CUDA is critical to Nvidia’s transformation into a software company. The parallel programming framework is behind software and service offerings in markets such as robotics, automotive and health care that can only run on Nvidia’s GPUs.

Nvidia is moving heavily towards Arm, the company behind its homegrown CPUs. CUDA already supports x86, but there is no RISC-V support in the roadmap.

CUDA's upcoming version 12, expected soon, has a number of optimizations for Nvidia's Arm-based CPUs called Grace. The chipmaker is pairing its latest hopper-based GPUs with Grace CPUs, which can communicate over a proprietary interconnect called NVLink that has five times the bandwidth of PCIe Gen 5, which will be used in x86 CPU and Nvidia GPU systems.

Nvidia's CEO Jensen Huang said in a press conference with Asia-Pacific media that Nvidia was an early adopter of RISC-V in its GPU controllers, but this is the best use of the architecture at present.

“We like RISC-V because it’s open source… but more importantly, it’s adaptable. We can use it for all sorts of interesting CPU configurations. However, RISC-V is not yet amenable to external third-party software, and it won’t be in the short term.”

Huang said that in contrast, the x86 and Arm architectures have a large software ecosystem that is not fragmented and is stable, regardless of which vendor it comes from.

The benefits of RISC-V being open source and adaptable also have drawbacks.

The RISC-V architecture is more like a chip version of Linux and can be freely licensed and modified. The goal is to let companies make their own chips at a low cost while reducing their reliance on proprietary x86 and Arm architectures, which must be purchased or licensed.

The RISC-V architecture has a base instruction set that companies can customize to put their proprietary extensions on top. For example, Nvidia competitor Imagination has produced its own RISC-V CPU, called Catapult, on which it can bundle its compatible graphics processors and AI. Imagination provides full software and debugging support. Similarly, other companies offer RISC-V AI chips with vector extensions and their own software stacks.

Therein lies the problem. Huang believes that this incoherent software ecosystem, with different software products for different chips, is a disadvantage of RISC-V. He points out that contributing to a fragmented ecosystem is not good for the development of RISC-V.

“In the long run, we’ll see how the world develops. But building a software-compatible ecosystem, an architecture-compatible ecosystem, that’s very, very hard,” Huang said, adding, “Can you do a RISC-V ecosystem like you did with Arm and x86? Sure, but that might take a decade or two.”

Huang's views may reflect Apple's views on RISC-V. Semiconductor analyst Dylan Patel said in a newsletter earlier this month that Apple will replace Arm controllers with RISC-V cores on non-user-facing parts. These parts are generally less dependent on system software.

RISC-V International, which drives the architecture and extension development, focuses primarily on hardware extensions. Open source developers and companies supporting RISC-V are developing and upgrading Linux 6.0 support for new extensions, as documented by Michael Larabel at Phoronix.

While software remains an issue, hardware adoption of the RISC-V architecture is growing. Intel is working with the Barcelona Supercomputing Center to develop RISC-V chips, and Google is working with SiFive to develop chips for artificial intelligence applications.

When asked about Huang's comments and Nvidia's stance on RISC-V CUDA, RISC-V International CEO Calista Redmond did not respond directly to the topic.

“We are seeing momentum and investment growing across the computing landscape, from the data center to mobile devices. The ecosystem is also evolving rapidly. Given the need for design flexibility, work is now well underway that might have taken decades given the introduction of a set of shared and open standards, such as our single hypervisor approach,” Redmond said in an email.