Comment on ROCm Is AMD’s No. 1 Priority, Exec Says

• just_another_person@lemmy.world ⁨1⁩ ⁨year⁩ ago

  The number of devices in use out in the world is a direct correlation to how useful a project like ROCm or CUDA is/could be. More devices means devs are more likely to utilize a specific language or library for a specific use. ROCm is open source and attempting to gain more ground simply by expanding to more devices which are already out there. My response to OP is just illustrating that fact.

  Example: Nvidia got an early foothold in the AI/ML game in the datacenter because they were first to platform traction with the CUDA toolkit and inference libraries. It’s horrible to use, but is useful. AMD is now trying to catch up to that by deploying alternative hardware and software that covers most of the same use-case, plus they now have APU and FPGA devices that Nvidia does not. That’s the tldr.

  source

  • poVoq@slrpnk.net ⁨1⁩ ⁨year⁩ ago

    Your comment doesn’t make sense. ROCm is a buggy mess that despite years of working on it AMD hasn’t been able to make work well at all.

    Intel’s oneAPI on the other hand is cross-vendor and by all appearances so far is good software that has a real shot at beating CUDA if AMD was not shooting itself in its own leg by riding the dead horse that is ROCm.

    source

    • just_another_person@lemmy.world ⁨1⁩ ⁨year⁩ ago

      I work with the entire CUDA toolkit on a daily basis, and it is also a mess. Nvidia is locked in though, and doesn’t plan any rework anytime soon (you can refer to their own statements on this). Any widespread alternative forces greater competition, and better products as a result.

      I’ve never met a single engineer who has worked on any of Intel’s acceleration toolchains, but they are just now getting new devices into the datacenter, so maybe it will gain in popularity.

      source
  • giacomo@lemm.ee ⁨1⁩ ⁨year⁩ ago

    lol, and that’s the argument OP was making; forget about ROCm and jump onboard with OneAPI

    source