Nvidia won

I mean, wtf is this?

Did Nvidia win?

Thalidomide Vintage Ad Shirt $22.14

Unattended Children Pitbull Club Shirt $21.68

Thalidomide Vintage Ad Shirt $22.14

  1. 3 weeks ago
    Anonymous

    >line goes up
    big if true, anyway I'm getting a 5090

  2. 3 weeks ago
    Anonymous

    lets see the fp16 numbers

    • 3 weeks ago
      Anonymous

      spbp
      kek. I actually read the pic and you're right. fricking sneaky Black folk.

  3. 3 weeks ago
    Anonymous

    This is a price chart

    • 3 weeks ago
      Anonymous

      thurpaburpa

    • 3 weeks ago
      Anonymous

      No, dumbass.
      It's wattage.

  4. 3 weeks ago
    Anonymous

    Nvidia driver woes are 100% on Gayland devs.
    >implicit sync
    Depreciated and nobody on any other system uses it. Explicit sync is superior. It should've been implemented by Gayland years ago but they decided to be stuck on OpenGL era principles for a decade. And yet nvidia added a quasi implicit sync support, probably to show how shit it is with flickering which btw also happens on AMD on mesa.
    >GBM
    EGLStream is superior and works better on nvidia. Despite all that, nvidia bent the knee and added unnecessary GBM support just to make those autists shut up.
    >Gayland has a history of not merging fixes from nvidia and instead pushing their own inferior ones to one up an nvidia engineer with more knowledge than all of their contributors combined.
    https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1120
    >Gayland devs are trying cut them some slack
    Look at any merges (even the one above), half the discussion is arguing about how to define certain things or uhhh let's cross reference mesa because we have no idea what we're doing.

  5. 3 weeks ago
    Anonymous

    80000 FP1 when

    • 3 weeks ago
      Anonymous

      infinite tflops fp0 when

  6. 3 weeks ago
    Anonymous

    >hardware developed faster than Moore's law
    "Haha see, it's real and we're winning!!!!1!1!111"
    >Hardware developed slower than Moore's law
    "So, Moore's law has been dead since 1991 anyways"

    • 3 weeks ago
      Anonymous

      Yes, that is how it works. People make statements in response to events as they unfold.
      This means that in opposite situations they may say opposite things.
      Maintaining one's prior position in the face of contrary evidence is the height of folly.

  7. 3 weeks ago
    Anonymous

    >flops
    go watch meatspin that many times

  8. 3 weeks ago
    Anonymous

    What's the energy consumption?

  9. 3 weeks ago
    Anonymous

    FP4????? Is this a joke? What's the range of FP4 anyways?

    • 3 weeks ago
      Anonymous

      enough for machine learning

    • 3 weeks ago
      Anonymous

      fake-AI shit only needs a single bit for "good" or "bad" scoring, some bits for the NN decision tree, and a little extra info that references the source and the output. The actual processing is wasted with full size FP, so 8- and 4-bit provides more calcs per second (at the expense of memory calls). But taining is so repetitive and failure prone, fake-AI shit doesn't suffer from memory access constraint problems because it needs to run thousands of iterations on Input A to successfully make Output A. The hardware can run thousands of trainings in the same generation using just one memory snapshot of the input data. The rest of the memory requirements are effectively network-internal data shuffling of the NN interacting with itself.

      Thus fp4 = multiples faster results over fp32. It's theoretically more than 8x faster because when you pack multiple small ops inside an fp32 cycle there is scheduling and other overhead in splitting the full fp32 back put to the smaller results.

  10. 3 weeks ago
    Anonymous

    >fp16
    >fp8
    >fp4

    • 3 weeks ago
      Anonymous

      FP4????? Is this a joke? What's the range of FP4 anyways?

      >I mean, wtf is this?
      >Did Nvidia win?

      can you read a chart?

      they're halving the precision to boost their numbers.

      >2026: 100,000 TFLOPS FP1.5

      This chart is enough to fool investors.

      • 3 weeks ago
        Anonymous

        seems its enough to fool most of IQfy too

        • 3 weeks ago
          Anonymous

          Why are you acting like that's a shocking revelation?

  11. 3 weeks ago
    Anonymous

    Won? Nvidia won years ago

  12. 3 weeks ago
    Anonymous

    More FLOPs
    Accelerate

  13. 3 weeks ago
    Anonymous

    fp4 being entirely useless outside things like ML that only care about the range of 0 to 1 is a flag.

  14. 3 weeks ago
    Anonymous

    blacked well

  15. 3 weeks ago
    Anonymous

    Moore's law isn't related to FLOPs you gaytard.
    It's the number of transistors that can fit on a chip, and we're brushing up against quantum physics when we make shit too small

    • 3 weeks ago
      Anonymous

      Nvidia man said that moore's law is AI flops per second so now it is. Cope and Seethe

  16. 3 weeks ago
    Anonymous

    Not with how price and power usage is.
    I'll probably need to install an 30 amp to run an video card for my computer

    • 3 weeks ago
      Anonymous

      Say hi to your new power cord anon

  17. 3 weeks ago
    Anonymous

    why are floating points going down?

    • 3 weeks ago
      Anonymous

      Because it's targetted at training gigantic AI models and that's what they use

    • 3 weeks ago
      Anonymous

      Because it's targetted at training gigantic AI models and that's what they use

      To add to this, the hardware in OP's image is made to be combined into datacenters with hundreds of exaflops (no exaggeration), and then it's used to train GPT5 in a matter of a few days instead of months/years. Obviously such a datacenter would cost billions too. This is not about gaming GPUs

  18. 3 weeks ago
    Anonymous

    Can't they fake the numbers by counting frame generation as real frames?
    like if 30 FPS = x TFLOPS, then 60 FPS with DLSS 3 = 2x TFLOPS.

  19. 3 weeks ago
    Anonymous

    Moore's law is about transistor's count, not FLOPS. Those two lines have nothing to do with one another.
    You're a colossal moron.

    • 3 weeks ago
      Anonymous

      >graph shows moore's law and flops
      >doesn't state either is invalid
      >shows predicted performance of both
      looks like you're the moron

  20. 3 weeks ago
    Anonymous

    Damn bro, I can't wait for the gtrx3dfx 60000 to have 1488 Exaflops of FP0.8.

  21. 3 weeks ago
    Anonymous

    Temporarily as always .
    Only real winner as always is timedaddy. Father Time wienervores all.

  22. 3 weeks ago
    Anonymous

    maybe they should have three green lines one for each FP

  23. 3 weeks ago
    Anonymous

    STRONG BUY
    STRONG BUY
    STRONG BUY

    (not fin adv)

  24. 3 weeks ago
    Anonymous

    >I mean

  25. 3 weeks ago
    Anonymous

    I fricking despise charts without scales on the axes

  26. 3 weeks ago
    Anonymous

    and infinite TFLOPS in 2026 with FP0 !!!

  27. 3 weeks ago
    Anonymous

    Show me the next graph when it arrives and I'll tell you how imbecile capital investors are and at what date the CEO will sell one wafer one die like only Elon Musk can con job.

  28. 3 weeks ago
    Anonymous

    >graph of predicted amount of transistors on an average CPU or something like that
    >a completely unrelated GPU graph that doesn't show CPU transistors, doesn't even show GPU transistors, doesn't even show fricking math operations per cycle, doesn't even fricking show fricking math operations * clock speed, it shows fricking FOUR BIT NUMBER OPERATIONS * clock speed

    Omg guys green line go up1!!1!11!!!!

  29. 3 weeks ago
    Anonymous

    holy frick who cares when is AMD or Nvidia going to make a GPU thats faster than a 6950XT for under $400.

    • 3 weeks ago
      Anonymous

      Hahahahaha! Never. They have zero incentive to pursue a market like that ever again. You can buy old stock and used cards, but nothing new, ever. Thats what chiplet design is all about. Up scaling or down scaling for production costs,’not consumer costs.

Your email address will not be published. Required fields are marked *