• DingoBilly@lemmy.world
    link
    fedilink
    English
    arrow-up
    17
    ·
    8 months ago

    My reading of chip wars and similar chip fabrication places is this stuff takes decades to develop and is insanely expensive. Is that not the case here?

    • AggressivelyPassive@feddit.de
      link
      fedilink
      English
      arrow-up
      14
      ·
      8 months ago

      They won’t start from scratch.

      You can license ARM cores, scale up the design a bit and get to first pre-production units within a year or two.

      OpenAI needs extremely large amounts of computing power and is almost 100% dependent on Nvidia to get it. That’s not a good situation to be in. Even if they can’t compete with Nvidia directly, building their own chips might be cheaper than paying Nvidia’s profit margin.

      • HeavyRaptor@lemmy.zip
        link
        fedilink
        English
        arrow-up
        2
        ·
        8 months ago

        I could kinda see them designing a chip maybe but that’s not what they are saying. Many companies have their own chip by going to AMD for example and using their designs and process nodes to create something more use-case specific. For example the steam deck’s soc is a custom design madefor Valve. Or as you said it is possible to licence and customise designs from Arm to create your own custom SOC. This is what smartphone manufacturers do. Then they take their designs to a foundry like TSMC to create actual chips out of them.

        But I don’t think that’s what openai is aiming for here, they want their own foundry. They themselves want to design and produce the chips.

        Not even considering the actual design of the chips, just being able to mass produce something on the scale of modern transistors with economical yields is an insane task.

        On some of the newest nodes, one aspect of lythography (the process of marking out transistors with a directed light) involves creating a micro-drop of certain molten metals. While the drop is falling, it is hit by a low power laser to change its shape to a ‘pancake’ like form. After this it is hit by a more powerful laser that vaporises the drop, crating a very specific wavelength of light. This light is then focused into an extremely sharp laser that marks out the transistor patterns on the target silicon. The process happens thousands of times a second.

        Of course you can’t just buy one of these machines on Amazon. Each company guards their newest process nodes with the utmost secrecy, as developing a new node can cost billions of dollars, not even mentioning what it costs to build, outfit, and run an actual foundry itself.

        This is one machine that is part of the complex process that is making microchips. Nvidia doesn’t even have their own foundry, they have been relying on Samsung and TSMC recently.

        Asking for trillions of dollars to build something from scratch is just so unbelievably silly. Almost as silly as someone actually giving it to them.

        • AggressivelyPassive@feddit.de
          link
          fedilink
          English
          arrow-up
          2
          ·
          8 months ago

          Building the entire pipeline initially would be insane, true.

          But as an aspiration, it’s not that dumb. Start with a highly optimized ARM chip in gen 1, add more ASIC components in gen 2, have some packaging in-house for gen 3, etc. etc.

          OpenAI seems to aim at the long game. Investing billions now to save trillions later could be sane.