11 comments

  • hassaanr 2 minutes ago
    While cool, quantization to FP4 is practically never lossless in actual use. A lot of providers are advertising high TPS on Kimi and GLM, but the models are functionally lobotomized and no longer close to frontier quality. Would love to see this not be true.
  • minraws 4 hours ago
    Can you folks add performance per watt as a metric to these comparisons, I honestly want to understand where AMD fits in the stack in terms of actual performance to dollars. I have had talks with companies wanting to build data centers outside of US and find it hard to source anything Nvidia in sufficient capacity and scale.

    If AMD is competitive performance per watt and roughly reliable in terms of software support which is what most folks outside of US prioritize above all else, since outside of China and US electricity tends to at a relative premium.

    Maybe if they make smaller data centers viable at the right price, AMD could be part of the stack outside of US where ever Nvidia is more limited in supply. Though I have genuinely no idea what sourcing an AMD GPU looks like.

    I have never seen a company use AMD outside of wafer and a couple others mostly in US.

    Genuinely intriguing or maybe not really (could be this stuff is common knowledge) and I am just stuck in my Nvidia bubble here.

    • kingstnap 2 hours ago
      A DGX B200 costs like ~$0.5 M and uses around 14 kW.

      If you plan to run it straight for 8 years 100% max usage thats around 1 GWhr.

      A gigawatt hour is a lot of energy but its not that much compared to the price of the actual machine. In Germany for example with its expensive energy thats about €100k worth, which spread over 8 years is pretty minor compared to the up front half mill.

      The real issue with high power consumption is not really the cost of energy but the limited powersupply you can get for a datacenter. A more efficient setup is highly desirable because it means you can fit more in the limited power hookup.

    • Twirrim 4 hours ago
      > I have never seen a company use AMD outside of wafer and a couple others mostly in US.

      There's a few using them, and even more starting to experiment with them. AMD has long been a source of disappointment around this side of things, so I'm hesitant to feel optimistic we'll finally get some competition. The market really needs viable competition to Nvidia, especially performance/watt.

    • craftkiller 4 hours ago
      • Schiendelman 3 hours ago
        It's not clear when this will be - AMD has slipped these dates likely to 2027.
    • latchkey 1 hour ago
      > I have never seen a company use AMD outside of wafer and a couple others mostly in US.

      Just because you haven't seen it doesn't mean it doesn't exist.

      We've serviced over 700 customers on our MI300x.

      • beffjezos 14 minutes ago
        700 customers on a direct to consumer API is not a lot.
    • technoabsurdist 3 hours ago
      AMD MI355X uses 1,400W per GPU and NVIDIA B200 uses 1,200W. So AMD uses about 16% more power.
      • vlovich123 3 hours ago
        Not how you measure performance per watt but generally it’s 20-60% worse at tok/s/watt not 16. It does have ~50% more memory (~100gb) which complicates the comparison.
  • nxtfari 1 hour ago
    I think we should make it illegal to not specify the quantization in the headline for these types of posts.
  • p1esk 2 hours ago
    There’s noticeable accuracy degradation when they switched from fp8 to mxfp4
    • throwdbaaway 2 hours ago
      And somehow they claimed that it is "lossless".
  • beffjezos 10 minutes ago
    This is very interesting and yet not at the same time. This looks to be optimized for single-stream LLM traffic which is not viable to serve in a production setting. It's only interesting to hobbyists that want to run the model locally.

    It's genuinely neat that AI can find the right optimization pathways in an AMD inference server to unlock this but at the same token (pun-intended) this is a classic case of benchmark hacking that doesn't stand up to real-world application.

    • wmf 7 minutes ago
      You got it backwards; it's ~200 on single stream so the 2,600 is achieved with ~13 streams.
    • technoabsurdist 5 minutes ago
      hi yes it’s not optimized for single stream it’s optimized for total node throughput
  • Schiendelman 3 hours ago
    I'm not surprised to see competition with Blackwell. Rubin is 5x faster than Blackwell at inference - Blackwell is the last generation Nvidia didn't optimize specifically for inference.

    If I'm missing something, please let me know!

    • boroboro4 17 minutes ago
      It's very unclear what's special in Rubin to be optimized for inference? I can see disaggregated bit (with having separate prefill and decoding nodes), but what else?
    • nullc 2 hours ago
      how do you get 5x faster at inference when inference is memory bandwidth limited? getting 5x the memory bandwidth of a h100 seems physically difficult.
      • Schiendelman 2 hours ago
        Rubin has 22TB/s of memory bandwidth vs Blackwell's 8TB/s. NVLink 6 doubles interconnect speed. Plus they're moving to 3nm from ~4nm.

        (Previously this comment said Rubin did native NVFP4, but Blackwell does too! Rubin just also trains with native NVFP4, which Blackwell does not.)

        • boredatoms 1 hour ago
          Moving to lower bits is not a slam dunk, the model itself might degrade too much
          • Schiendelman 1 hour ago
            Of course, but for most workflows it's fine.
        • zackangelo 1 hour ago
          Blackwell supports nvfp4 natively.
          • Schiendelman 1 hour ago
            You're right - Rubin is better at NVFP4 training, not inference, thank you for catching me!
            • boroboro4 18 minutes ago
              What does it mean it's better at nvfp4 training? What's different between training and inference to make this true?
  • AussieWog93 4 hours ago
    The 2600 tok/s is an "aggregate", not the actual throughput.
  • alienbaby 1 hour ago
    I'm interested if anyone knows how much legwork the assumed 60% cache hit, plus running a quantised model is doing? Esp. compared to what the headline half implies is a full fat GLM5.2
  • killingtime74 1 hour ago
    No word on what this actually means as a consumer. What's the price. Is it lower than NVIDIA serving?
  • oDot 4 hours ago
    Do these providers have 80+% gross margins or is something eating into them? Maybe utilization?
    • technoabsurdist 4 hours ago
      hi i work at wafer. no the margins are lower averaging at about ~40%. utilization is one of the highest order bits in determining margins here, yes.
  • yieldcrv 4 hours ago
    Agentic coding drivers for different architectures is a massive unlock for the world

    So much compute is under utilized waiting for a savant or company to prioritize an architecture, and now all the other engineers can tackle this at any time if they get inspired on the right prompts

    • technoabsurdist 3 hours ago
      this is exactly our thesis at wafer :) thank you for the support
    • yogthos 3 hours ago
      Personally, I can't wait till something like this starts getting to consumer level. https://www.anuragk.com/blog/posts/Taalas.html
      • yieldcrv 3 hours ago
        That’s pretty fascinating, Apple has some innocuous LLMs and transformers baked into its devices and leveraging their neural chipset

        So I could see something like this where the neural chipset has an LLM that cant be so easily updated baked into it, until you get a new device