Hot Chips 2026: CUDA Targets RISC-V – By Chester Lam
Comments
LogTrim
nxobject
Beyond the significant PCI-e requirements, I'm tickled pink that many of the server platform requirements are "do what x86 does"... ACPI, UEFI, a TPM. I hope they do a better job with standards compliance than some of the shadier Wintel OEMs!
classichasclass
And Nvidia is big enough that they could impose some order on the ecosystem that the ecosystem so far has either been unable or unwilling to impose on itself. The wild west of one-offs can continue to make their application-specific designs while everyone else will presumably benefit from a reasonably standardized platform.
nozzlegear
All these AI do is eat hot chip and lie
carabiner
Any fab born after 2005
rldjbpin
very ironic but easily explainable. as they sell their own integrated platform now, with server rack partners behaving like graphics cards oems now.
to be able to maximize customized solutions, they would like to own the entire stack which is not already commodified. the cpu remains the only final frontier in this.
there are always side-effects possible, but i wonder if all these efforts can be applied outside of ai ever.
boredatoms
When does bigsky ship?
physicsguy
CUDA used to target POWER9 architecture but it was so niche that you ended up compiling from source literally everything since no wheels were available for it. EasyBuild and Spack were full of half broken recipes for the magic incantations necessary to build fundamental packages for HPC like TensorFlow, etc.
einpoklum
CUDA is proprietary technology, guarded by patents and mostly closed-source. And its originator, NVIDIA, while being a member of the Khronos consortium, and technically supporting OpenCL, makes it much more difficult for you to do anything in OpenCL than in CUDA - even though it is sometimes almosy zero effort to enable some OpenCL equivalent (e.g. : Half-precision types. They exist in CUDA; in OpenCL on NVIDIA - you can't have them).
So, live with CUDA on other hardware? No thank you.
djmips
hot chips. hot planet.
varispeed
Just make more RAM.
goglidesdev
[flagged]
bizstrategy10
[dead]
Ozzie-D
[flagged]
The interesting part is that this is less "CUDA now runs on RISC-V" and more "Nvidia is defining what a CUDA-capable RISC-V server has to look like."
That could end up being a pretty influential de facto server profile for the architecture.