• bioemerl@kbin.social
    link
    fedilink
    arrow-up
    21
    arrow-down
    2
    ·
    1 year ago

    As it fucking should be, this is like the third time they have tried to redesign the chip to get around export restrictions

    • baconisaveg@lemmy.ca
      link
      fedilink
      arrow-up
      11
      arrow-down
      5
      ·
      edit-2
      1 year ago

      And yet all of the good, open source LLM work is coming out of China. But hey, maybe this will spur some ROCm development so AMD cards don’t suck so fucking hard in that space.

        • baconisaveg@lemmy.ca
          link
          fedilink
          arrow-up
          4
          ·
          1 year ago

          I liked your original comment before the edit, it’s true. AMD only has consumer grade hardware, which is the main reason the majority of the AI projects don’t support it. Code written for CUDA can work on a A100 just as well as a 3090, there’s no money in developing software to run on someone’s 4 year old gaming PC.

      • fruitycoder
        link
        fedilink
        arrow-up
        3
        ·
        1 year ago

        Right? It’s really unfortunate that so many of the Chinese people that make awesome stuff get limited because CCP has garnered so much distrust that they won’t use AI to advance their surveillance and authoritarian control systems both domestically and abroad.

        I mean it was also a shame how much the actions of the NSA weakend trust in West as well.

  • tal@lemmy.today
    link
    fedilink
    arrow-up
    18
    arrow-down
    1
    ·
    edit-2
    1 year ago

    I’m deeply skeptical that export controls are going to work with China if the item in question is a consumer GPU unless you’re going to radically redesign the things to do something extremely obnoxious like require cryptographically-secure online activation every boot. There are just too many individuals and small businesses in random other countries who are going to be able to buy them and resell them. The chain over which a GPU travels is too far disconnected from entities that we can directly affect.

    Maybe they can induce a very limited delay, or increase costs to a limited degree, but I assume that that’s not what the goal is.

    You can only effectively export-control something if, all along the chain from the manufacturer to the end user, you can clearly distinguish between a legit and non-legit user. There’s no way that some random electronics store in Kazakhstan is going to know whether the person buying a 4090 with cash is going to sell it in China or use it themselves.

    • baconisaveg@lemmy.ca
      link
      fedilink
      arrow-up
      10
      ·
      edit-2
      1 year ago

      It’s also kind of ridiculous to think that anyone using a consumer grade GPU is a threat. No one doing work with AI is using a 4090, they’re using enterprise grade cards with 80GB of vram (H100/A100, even P40’s in a pinch). The 4090 is a nice toy for running Stable Diffusion and low grade ERP LLM models and that’s about it.

    • MonkCanatella
      link
      fedilink
      arrow-up
      4
      ·
      1 year ago

      Bad logic. If there’s risk involved and they can’t get the. Legitimately, it’s going to add a significant overhead.