3 ms·
Hot Chips 2026: CUDA Targets RISC-V
- djmips 2mo agohot chips. hot planet.
- nozzlegear 2mo agoAll these AI do is eat hot chip and lie
- carabiner 2mo agoAny fab born after 2005
- camel-cdr 2mo agorelated: https://www.sifive.com/development-platforms/sifive-bigsky-sf-2u870-datacenter https://www.sifive.com/development-platforms/sifive-bigsky-s...
- boredatoms 2mo agoWhen does bigsky ship?
- LogTrim 2mo agoThe 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.
- nxobject 2mo agoBeyond 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 2mo agoAnd 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.
- varispeed 2mo agoJust make more RAM.
- einpoklum 2mo agoCUDA 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.
- bizstrategy10 2mo ago[dead]
- Ozzie-D 2mo ago[flagged]
- goglidesdev 2mo ago[flagged]
- physicsguy 2mo agoCUDA 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.
- rldjbpin 2mo agovery 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.
- DiabloD3 2mo agoDon't really know why this is interesting, CUDA is kind of a niche legacy API that Nvidia has kept on life support a little too long. Just use Vulkan for compute.
- brucehoult 2mo agoOr Mojo, recently acquired by Qualcomm, and even more recently fully open-sourced under Apache 2 license.