AMD talks 1.2 million GPU AI supercomputer to compete with Nvidia — 30X more GPUs than world's fastest supercomputer

AMD
(Image credit: AMD)

Demand for more computing power in the data center is growing at a staggering pace, and AMD has revealed that it has had serious inquiries to build single AI clusters packing a whopping 1.2 million GPUs or more.

AMD's admission comes from a lengthy discussion The Next Platform had with Forrest Norrod, AMD's EVP and GM of the Datacenter Solutions Group, about the future of AMD in the data center. One of the most eye-opening responses was about the biggest AI training cluster that someone is seriously considering.

When asked if the company has fielded inquiries for clusters as large as 1.2 million GPUs, Forrest replied that the assessment was virtually spot on.

Morgan: What’s the biggest AI training cluster that somebody is serious about – you don’t have to name names. Has somebody come to you and said with MI500, I need 1.2 million GPUs or whatever.

Forrest Norrod: It’s in that range? Yes.

Morgan: You can’t just say “it’s in that range.” What’s the biggest actual number?

Forrest Norrod: I am dead serious, it is in that range.

Morgan: For one machine.

Forrest Norrod: Yes, I’m talking about one machine.

Morgan: It boggles the mind a little bit, you know?

1.2 million GPUs is an absurd number (mind-boggling, as Forest quips later in the interview). AI-training clusters are often built with a few thousand GPUs connected via a high-speed interconnect across several server racks or less. By contrast, creating an AI cluster with 1.2 million GPUs seems virtually impossible.

We can only imagine the pitfalls someone will need to overcome to try and build an AI cluster with over a million GPUs, but latency, power, and the inevitability of hardware failures are a few factors that immediately come to mind.

AI workloads are extremely sensitive to latency, particularly tail latency and outliers, wherein certain data transfers take much longer than others and disrupt the workload. Additionally, today's supercomputers have to mitigate the GPU or other hardware failures that, at their scale, occur every few hours. Those issues would become far more pronounced when scaling to 30X the size of today's largest known clusters. And that's before we even touch on the nuclear power plant-sized power delivery required for such an audacious goal.

Even the most powerful supercomputers in the world don't scale to millions of GPUs. For instance, the fastest operational supercomputer right now, Frontier, "only" has 37,888 GPUs.

The goal of million-GPU clusters speaks to the seriousness of the AI race that is molding the 2020s. If it is in the realm of possibility, someone will try to do it if it means greater AI processing power. Forest didn't say which organization is considering building a system of this scale but did mention that "very sober people" are contemplating spending tens to hundreds of billions of dollars on AI training clusters (which is why millions of GPU clusters are being considered at all).

Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

  • A Stoner
    And even with that, they would not even have the processing power of an insect at their disposal. We still have no factual intelligence from all the AI spending that has happened. No AI knows anything at all as of yet.
    Reply
  • JRStern
    Well Musk was just out there raising money for 300,000 GPUs, we're talking billions or trillions before they're all installed and usable, not to mention gigawatts of power to run. OTOH this is crazy stuff, IMHO, and perhaps Elon isn't hip to the news that much smaller LLMs are now being seen as workable so maybe nobody will need a single training system with more than 300 or perhaps 3000 GPUs, to do a retrain within 24 hours. And maybe whole-hog retrains won't be as necessary anymore, either.

    So AMD is just trolling, is what this comes down to, unlikely to actually build it out.
    Reply
  • Pierce2623
    The record Dynex set recently was only a quantum record and the record they beat wasn’t even real quantum computing. The record they beat only involves 896 GPUs
    Reply
  • jeremyj_83
    It literally said "For instance, the fastest operational supercomputer right now, Frontier, "only" has 37,888 GPUs." in the article. Frontier has 1.1 exaFLOPs of computing power just so you know.
    Reply
  • DS426
    Usually business is all about ROI and profit but... really, c'mon, someone show me the math on how investments like this pay off without losing money?? We're also talking about cooling, electric bills, sys admins, and so on, so... wtf is so magically about a (relatively?) well-trained and advanced AI LLM or such that costifies this?

    Seriously, not being a hater just to hate but again being on the business side of things in IT, I need to see some math.

    On another note, at least some folks are seeing the value in not paying ridiculous cash just to have "the best" (nVidia) whereas AMD can honestly and probably provide a better return on investment. Kind of that age-old name brand vs. generic argument.

    Still mindblown over here. How many supercomputers have more than 1.2 million CPU's? I know this doesn't account for core counts but holy smokes, we're clearly not talking apples to apples here!! Pretty sure a mini power plant is literally needed to sit beside a datacenter/supercomputing facility like this.
    Reply
  • oofdragon
    I honestly don't get it. Ok so someone like Elon is considering 300 thousand GPUs like Blackwell's, spending in the order of billions just to buy them, then you have the electric bill and maintenance as well every month. In what war can he possible make a profit out of this situation?
    Reply
  • abufrejoval
    Nice to see you reading TNP: it's really one of the best sites out there and on my daily reading list.

    And so are the vultures next door :-) (the register)
    Reply
  • ThomasKinsley
    Not to get all cynical, but this sounds like a bit of a stretch to me. The reporter gave the random number 1.2 million and the AMD staff member responded with, "It’s in that range? Yes." A range needs more than one number. Are we talking 700,000? 1 million? 1.4 million? There's no way to know.
    Reply
  • kjfatl
    If Musk is serious about the 300,000 GPU's it makes perfect sense that the design would support an upgrade path where compute modules could be replaced with future modules with 2X or 4X the capacity.
    The most obvious use for such a machine is for constant updates to self-driving vehicle software. Daily or even by the minute upgrades are needed for this to be seamless. This is little different than what Google or Garman does with maps. When 'interesting' data is seen by vehicles it would be sent to the compute farm for processing. Real-time data from a landslide just before the driver ran off the side of the road would qualify as 'interesting'. Preventing the crash in the next landslide would be the goal.

    This sort of system is large enough to justify custom compute silicon supporting a limited set of models. This alone might cut the hardware requirements by a factor of 4. Moving to Intel 14A or the equivalent from TSMC or Samsung might give another factor of 8 toward density. Advanced packaging techniques might double it again. Combining all of these could provide a machine with the same footprint and power envelope of today's supercomputer with 30,000 GPUs.
    Reply
  • shawman123
    How much power would Million GPUs would consume. its seems off the charts if all of them are fully used. !!!
    Reply