Sidestepping GPU ban, Chinese factories dismantle and transform Nvidia RTX 4090 gaming cards into AI accelerators - eviltoast
  • Riskable@programming.dev
    link
    fedilink
    English
    arrow-up
    81
    ·
    edit-2
    1 year ago

    This article is misleading: The cards are already “AI accelerators” they just come in monstrous sizes and coolers that aren’t suitable for cramming more than one into a computer (server) case.

    What the Chinese chop shops are doing is removing the components and resoldering them on to smaller PCBs and putting on smaller (but more powerful and jet-engine loud) exhaust-style coolers.

    Basically it’s just old fashioned harvesting and re-using of PCB components. A common thing during the shortages of the COVID-19 pandemic.

    • MonkderZweite@feddit.ch
      link
      fedilink
      English
      arrow-up
      6
      ·
      edit-2
      1 year ago

      Btw, about the

      smaller (but more powerful and jet-engine loud) exhaust-style coolers.

      Why not make double-sided racks with fans twice the size, higher cooling efficiency?

      • Player2@sopuli.xyz
        link
        fedilink
        English
        arrow-up
        16
        ·
        1 year ago

        Not sure exactly what you’re asking, but usually servers are built in a way that several units can be stacked on top of each other in a rack, which makes it important that air comes in from the front and exits from the back. The smaller coolers then allow you to cram more GPUs next to each other since the higher heat density won’t matter due to the airflow speed inside the case.

    • teft@startrek.website
      link
      fedilink
      English
      arrow-up
      28
      arrow-down
      1
      ·
      edit-2
      1 year ago

      Maybe this is why it was so easy to get a PS2 in iraq. That’s crazy, 20 years later and mystery solved for me.

    • Never_Sm1le@lemdro.id
      link
      fedilink
      English
      arrow-up
      12
      arrow-down
      1
      ·
      1 year ago

      There’s also the one with ps3, which is more suitable because ps3’s cpu is a supercomputer one

      • WHYAREWEALLCAPS@kbin.social
        link
        fedilink
        arrow-up
        13
        ·
        1 year ago

        It made sense. There were also university CS departments doing the same thing. The Cell processor in the PS3 was pretty damn powerful for the price at the time.

        • sir_reginald@lemmy.world
          link
          fedilink
          English
          arrow-up
          12
          ·
          1 year ago

          and iirc, it had Linux support, so it was perfect because Sony was selling the hardware at a loss to sell more units and consequently sell more games.

          when people started buying ps3s to install Linux, Sony made a firmware update to block this.

  • AutoTL;DR@lemmings.worldB
    link
    fedilink
    English
    arrow-up
    8
    ·
    1 year ago

    This is the best summary I could come up with:


    One user from the Tieba Baidu forums recently visited a fascinating Chinese factory dedicated to the graphics card recycling business.

    The workers diligently disassembled the custom GeForce RTX 4090 graphics cards, extracted the AD102 silicon and GDDR6X memory modules, and repurposed them onto a “special reference” PCB with a blower-style cooler.

    These factories aim to reduce the custom GeForce RTX 4090’s footprint with a more compact blower style to turn them into “AI accelerators.”

    The typical blower only takes up two PCI slots, allowing you to stick four or more GeForce RTX 4090 graphics cards into a rack or clusters in the case of the more prominent players, like the big AI farms.

    One set of leaked photographs show a few pallets with a boatload of GeForce RTX 4090 graphics cards from vendors like Gainward and Palit.

    The GeForce RTX 4090’s soaring pricing should be a temporary phenomenon as Nvidia had reportedly rushed as many units as the chipmaker could to its Chinese customers before the sanctions took effect.


    The original article contains 611 words, the summary contains 169 words. Saved 72%. I’m a bot and I’m open source!