Instinct MI300 Could be Fastest Product to $1 Billion in AMD History

News
By
published

AMD pins a lot of hopes on Instinct MI300-series products.

AMD MI300X
(Image credit: AMD)

For years AMD has been trying to grab a share of the datacenter GPU market from Nvidia. It hasn't had much success, largely due to poor software support for its Instinct GPUs. But it looks like the situation is improving and the company finally has its breakthrough product. At least, the chief executive of AMD believes that the upcoming Instinct MI300-series will be the company's fastest product to ramp to $1 billion sales.

"We now expect datacenter GPU revenue to be approximately $400 million in the fourth quarter and exceed $2 billion in 2024 as revenue ramps throughout the year," said Lisa Su, chief executive of AMD, at the company's earnings call with analysts and investors (via Motley Fool). "This growth would make MI300 the fastest product to ramp to $1 billion in sales in AMD history." 

There is a reason why AMD's Instinct MI300-series is expected to be considerably more successful than its predecessor. The new product family will address not only supercomputers and select datacenters, but also cloud service providers (CSPs) which plan to use these processors for AI training and inference as they find their performance and software capabilities competitive.

In fact, AMD is already shipping Instinct MI300A accelerated processing units to the Lawrence Livermore National Laboratory for the El Capitan supercomputer, one of the first machines with performance of over 2 ExaFLOPS. The AMD Instinct MI300A is a multi-chiplet design featuring three eight-core Zen 4 chiplets and multiple CDNA3 chiplets for artificial intelligence (AI) and high performance computing (HPC) workloads. 

In the coming weeks the company intends to start shipments of its Instinct MI300X processor to cloud service providers. Unlike MI300A, AMD's MI300X solely uses CDNA3-based chiplets for AI and HPC and acts like a typical compute GPU. 

"Production shipments of Instinct MI300A APUs started earlier this month to support the El Capitan Exascale supercomputer, and we are on track to begin production shipments of Instinct MI300X GPU accelerators to lead cloud and OEM customers in the coming weeks," said Su. 

On the hardware front, the development and validation of AMD's Instinct MI300A and MI300X accelerators are progressing as planned, with performance now either meeting or surpassing the company's internal expectations, according to the executive. As far as software is concerned, AMD has broadened its AI software ecosystem, achieving significant advancements in enhancing the performance and features of its ROCm platform this past quarter. Besides, ROCm has been integrated into the mainstream PyTorch and TensorFlow ecosystems. Furthermore, Hugging Face models are now consistently updated and validated to run on AMD hardware in general and Instinct accelerators in particular.

Anton Shilov
Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

18 Comments Comment from the forums
  • Co BIY
    No knock on AMD or their new product but ...

    A Billion just isn't what it used to be.
    Reply
  • redgarl
    Co BIY said:
    No knock on AMD or their new product but ...

    A Billion just isn't what it used to be.
    You don`t get it... Lisa said it is the most successful product AMD produce to date...

    "This growth would make MI300 the fastest product to ramp to $1 billion in sales in AMD history."
    Reply
  • Giroro
    So is there suddenly, like, a lot more demand for GPUs with no tensor cores that can't run CUDA?

    Because it doesn't sound like AMD has solved either of those problems.
    Reply
  • Vanderlindemedia
    If your in stocks, this is i think the time to start spending those on AMD.
    Reply
  • The Hardcard
    Giroro said:
    So is there suddenly, like, a lot more demand for GPUs with no tensor cores that can't run CUDA?

    Because it doesn't sound like AMD has solved either of those problems.
    Read more on AMD Instinct. The previous generation had matrix math cores and the MI 250X can dance with the A100 in most use cases.

    ROCm HIP stack has functional parity with CUDA. Worst case you have to modify your code some, but a lot of code will run unmodified. Sometimes you can even leave the CUDA headers intact.

    AMD’s remaining problem is that ROCm is still harder and more confusing to use. The documentation is behind and there are more code specific, GPU specific quirks and issues. But those gaps are getting closed.
    Reply
  • Avro Arrow
    I was wondering when El Captian would be coming online because I think it was like three years ago when Cray was first talking about it. I remember how they wanted to use Radeon Instinct because of the great successes that the US government already had with the Frontier supercomputer. The capabilities of El Capitan sounded pretty exciting but then we heard nothing about for a long time. Supercomputers are the one topic that actually makes me somewhat interested in what is possible on the server side.
    Reply
  • Order 66
    Avro Arrow said:
    I was wondering when El Captian would be coming online because I think it was like three years ago when Cray was first talking about it. I remember how they wanted to use Radeon Instinct because of the great successes that the US government already had with the Frontier supercomputer. The capabilities of El Capitan sounded pretty exciting but then we heard nothing about for a long time. Supercomputers are the one topic that actually makes me somewhat interested in what is possible on the server side.
    I know that supercomputers are not meant for gaming. I would like to see a supercomputer that is made for gaming, or maybe a game being coded well to scale well on a supercomputer. I wonder if games were coded to run on supercomputers or if supercomputers were designed for gaming, what the performance would be. I bet cyberpunk could be run at 8k max settings at 120 fps, if it scaled well and was coded properly, or is this totally impossible to do?
    Reply
  • Avro Arrow
    Giroro said:
    So is there suddenly, like, a lot more demand for GPUs with no tensor cores that can't run CUDA?

    Because it doesn't sound like AMD has solved either of those problems.
    That's because you think that the entire computing world is PCs when they're actually a tiny fraction of what is actually out there. The server/data centre side makes the PC market look like a bad joke in comparison. What, do you think that these supercomputers are being used for Blender? Man, that's just hilarious!

    Unfortunately for nVidia, computers of this scale are used for simulating extremely complex atomic and molecular interactions like protein folding and nuclear chain reactions. WTH is CUDA going to do for that? Literally nothing, as this list of fastest supercomputers in the world shows us.

    The top-5 fastest supercomputers in the world are the following:

    1) Cray Frontier (AMD EPYC, Radeon Instinct), USA
    2) Fujistu Fugaku (Fujitsu AFX64 ARM, no dedicated GPU cores), Japan
    3) Cray LUMI (AMD EPYC, Radeon Instinct), Finland
    4) Atos Leonardo (Intel Xeon, nVidia Ampere), Italy
    5) IBM Summit (IBM POWER9, nVidia Tesla), USA

    When El Capitan comes online, three of the top-5 (#1, #2 & #4) supercomputers in the world will be using Radeon GPUs with the fastest supercomputer using nVidia GPUs will be #5.

    If CUDA actually made a difference in supercomputers, we wouldn't see Radeons in the fastest and most expensive supercomputers in the world, would we?

    So, with almost all of the fastest supercomputers in the world using Radeons, just what "problems" are you referring to?
    Reply
  • Avro Arrow
    Order 66 said:
    I know that supercomputers are not meant for gaming. I would like to see a supercomputer that is made for gaming, or maybe a game being coded well to scale well on a supercomputer. I wonder if games were coded to run on supercomputers or if supercomputers were designed for gaming, what the performance would be. I bet cyberpunk could be run at 8k max settings at 120 fps, if it scaled well and was coded properly, or is this totally impossible to do?
    That will never happen because games don't need a lot of cores to play so what is a computer with hundreds of thousands of cores going to do with a game? Even the twelve and sixteen-core Ryzen 9 X3D CPUs are showing issues with CCX selection, making them slower at gaming than the single CCX R7-7800X3D.

    Here's what a fast PC is like:
    Here's what a fast supercomputer is like:
    You really can't use one for the purposes of the other because they're completely different animals. Trying to use a supercomputer for gaming would be like trying to win a race with a Tar Sands Earth Mover. It's just not going to happen. ;)
    Reply
  • Order 66
    Avro Arrow said:
    Even the twelve and sixteen-core Ryzen 9 X3D CPUs are showing issues with CCX selection, making them slower at gaming than the single CCX R7-7800X3D.
    What do you mean? I have heard nothing about it.
    Reply