Alternative(s) to run CUDA on non-Nvidia hardware

Can anyone finally free AI coders from Nvidia, or is this another doomed escape plan

TLDR: Spectral Compute says it can make Nvidia-only AI code run on other chips, which could loosen one of tech’s biggest lock-ins. Commenters love the idea but many are openly cynical, joking that these projects usually start bold and end abandoned.

A small London startup just walked into one of tech’s most locked-up kingdoms and basically said: what if Nvidia’s software didn’t have to stay married to Nvidia’s chips? Spectral Compute says its tool, SCALE, can take code written for Nvidia’s world and run it on rival hardware like AMD, while even squeezing more speed out of Nvidia’s own cards. That’s a huge deal because this software language has become the default way many scientists and AI developers tell powerful graphics chips what to do.

But the real show was in the comments, where the crowd split into skeptics, escape artists, and battle-scarred veterans. One camp rolled its eyes hard at yet another “CUDA alternative,” with one commenter joking that these projects always end the same way: a flashy launch, support for "3 operations," and then a ghost-town Discord asking, “any updates?” Others jumped in with rival routes: ZLUDA got name-dropped as the scrappy open-source option, while another commenter said the cleanest move is to ditch the whole Nvidia ecosystem and go with Vulkan instead, essentially yelling, burn the moat down.

There was also classic niche-tech chaos: one developer popped up to say they’re trying to make CUDA run on Apple Macs through cuda-metal, which feels very much like the internet refusing to let a bad idea die until it becomes genius. The strongest vibe? Hope mixed with deep, deep trust issues. People want freedom from Nvidia’s grip, but the comments read like a support group for folks who’ve been disappointed before.

Key Points

  • Spectral Compute is developing SCALE, a compiler-based drop-in replacement for Nvidia’s NVCC to run CUDA code on non-Nvidia hardware.
  • The startup was founded in 2018 by four engineers who were frustrated by Nvidia GPU costs, hardware lock-in, and weak performance from alternative compilers.
  • SCALE uses Clang and LLVM, initially targeted AMD GPUs, and is expanding toward other AI accelerators while also supporting Nvidia GPUs.
  • The article compares SCALE with HIPIFY, SYCLomatic, and ZLUDA, noting tradeoffs in PTX support, code coverage, manual migration effort, and performance.
  • Spectral says its published benchmarks show nearly a 6x performance boost on AMD GPUs compared with HIPIFY-based conversion to AMD’s ROCm stack, and that it validates output numerically against NVCC.

Hottest takes

"No reason to tie yourself to Nvidia's moat" — DiabloD3
"every CUDA alternative follows the same arc" — luciana1u
"works for 3 operations" — luciana1u
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.