Elektrine lite

← Feed

@lina@vt.social

2026-04-30 05:11 UTC

@ben TBF it's about one full 42U rack, going by the numbers (photo shows half of it, looks like 10U per chassis and there's 4), so about one (large) fridge worth of servers (and many, many times the power consumption of two office towers, I'm guessing easily 30kW which is probably why they split it into two racks and the photo only shows one, can only use so much power per rack... so it's two large fridges in practice). Still nothing compared to (OpenAI style) LLM training of course ^^

Replies (1)

  • @ben@mastodon.lubar.me 2026-04-30 12:24

    @lina my "two office computer towers" comparison was mainly about physical footprint size. I don't know the specifics of what hardware giant commercial LLMs like ChatGPT run on, but I would assume that the type of hardware is similar because the math being done is basically the same for any deep learning system (it's weighted sums all the way down). In the video, John McDonald talks about how these servers are only used for the inferencing step. The training is (or at least was at the time of this video) done locally on his work computer.

    Open ##1875342