Image: AMD

It appears that AMD has at least one innovative answer to NVIDIA’s upcoming multi-chip-module (MCM) Hopper graphics architecture. As spotted by various users on social media, red team has filed a patent for a radical new GPU design that leverages a system comprising chiplets for what could be a massive leap in graphics performance. The technology would presumably be used in the company’s future Radeon RX graphics cards.

“A chiplet system includes a central processing unit (CPU) communicably coupled to a first GPU chiplet of a GPU chiplet array,” an abstract reads. “The GPU chiplet array includes the first GPU chiplet communicably coupled to the CPU via a bus and a second GPU chiplet communicably coupled to the first GPU chiplet via a passive crosslink. The passive crosslink is a passive interposer die dedicated for inter-chiplet communications and partitions systems-on-a-chip (SoC) functionality into smaller functional chiplet groupings.”

AMD points out that the development of a chiplet GPU has been hindered by various challenges, such as figuring out how to distribute work properly across so many parts using traditional programming models, but the company believes that it may have a solution with the use of high bandwidth passive crosslinks for coupling chiplets.

Join the Conversation

12 Comments

  1. Curious how they are going to do this from a software-perspective. It will either require a ton of driver development (per game/engine) or they need to make multi-gpu a base part of Vulkan and DX before this type of GPU releases.

    It’s likely way too much work ($$$) for a per game/engine implementation.

  2. Curious how they are going to do this from a software-perspective. It will either require a ton of driver development (per game/engine) or they need to make multi-gpu a base part of Vulkan and DX before this type of GPU releases.

    It’s likely way too much work ($$$) for a per game/engine implementation.

    There may need to be some tuning, and it’ll probably be a combination of driver work alongside some API flags exposed to software.

    However, current software should be able to run well enough with just drivers, if AMD is pushing this now.

    And of course, the big deal here is that chiplets would mean that both performance has the potential to scale up as more chiplets are added, while price has the potential to scale down given how cheap each chiplet is to produce individually.

  3. They’ll probably have to create a smart interposer of sorts if thats even a thing?

    That’s essentially what ‘chiplets’ implies, as I’ve read about them so far, but that may not be necessary or even the case. If they don’t use an interposer the interconnects between the chiplets will definitely be involved.

  4. Curious how they are going to do this from a software-perspective. It will either require a ton of driver development (per game/engine) or they need to make multi-gpu a base part of Vulkan and DX before this type of GPU releases.

    It’s likely way too much work ($$$) for a per game/engine implementation.

    Software is something AMD need to work on, a lot, so this could be problematic…

  5. I don’t think this will be a huge issue. GPUs are already highly parallel – just need to nail down the comms between modules, and AMD has been working on that for a while with Infinity Fabric.

    At least in the beginning, I think this approach will only be used where monolithic would just be too big to have reasonable yields — so very high end only ( at least at first).

    Seems like it would be an advanced form of SLI/CF – the comms chip would have to handle all of the comms that were previously handled via the driver — that makes it transparent to devs. And since everything is on the same package rather than over a bus, you can keep comms over a fabric at pretty high speed. It’s not exactly SLI/CF, but highly optimized on a hardware / firmware basis, and not entirely something we haven’t seen in the past.

    It won’t be 100% scaling, which is why on a per core/ SP basis monolithic will win, but MCM will be much more cost competitive and scale to higher core counts.

    The three keys to success are transparent to devs, cost per core, and total package performance. If you can nail all of those, it will take off wildly.

    As that comms chiplet gets optimized over time (probably generations), will see MCM start to trickle down into lower tiers as the cost / performance factor starts to kick in.

  6. It’s not exactly SLI/CF

    We should probably get this comparison out of the way: it’s not even close, with the obvious exception that there’s more than one die involved.

    The bandwidth between dies will be an order of magnitude higher, and they’ll be cooperatively rendering, not competing :)

  7. It’s more akin to what they are doing with CPU’s and infinity fabric. In a GPU multiple chiplets shouldn’t appears as anything other than a single GPU package. Not even close to CF/SLI.

    I’m kind of surprised that 6000 series didn’t use this.

  8. Just speaking sillyness for fun, because I know nothing, but was thinking like an interposer that has circuitry to like hold a data/ frame buffer to/ from each chiplet, then just puts it in order of sequence and thats what goes out in total. I think this will forcibly make the frame rate fixed at x number, as each chiplet would have to spitout a frame or partial frame no matter what at the ordained sequence and I think the sequence would have to be fixed to avoid having to have the chips talk to each other back and forth….
    I think this would make the interposer like a distributed io die of sorts.
  9. What’s the timeframe on this?
    If Nvidia has omething similar in the works, and we might see it in the next 2 years or so…..?

    I’m thinking a $800 10Gb 3080 would last 2 years nicely and then it’s new tech time…
    (well of course id have to be able to buy something first)

  10. If they have solved the interconnect latencies, then this is excellent news.

    GPU’s are a much better application for chiplets than CPU’s be3cause they scale almost perfectly with added shader units/cores/whatever, and modern GPU’s are getting really large, so yields must be a huge issue.

    Imagine just being able to tack on more chiplets if you need more performance.

    This could be real good.

Leave a comment