AMD announces open-sourcing of its GPU software stack and documentation, including MES

AMD Radeon RX 7900 XTX
(Image credit: YouTube - AMD)

Following pushes from the community to do so— particularly from George Hotz' Tiny Corp— AMD has finally started open sourcing some of their GPU software stack and documentation, including open-sourcing their MES (Micro-Engine Scheduler). Open-sourcing of the MES was one of the things that Tiny Corp predicted AMD would do last month, though they seemed to have predicted a turnaround time well sooner than the one we actually got. As-is, we're still waiting for the full release— this is just AMD announcing their intent to do it.

For those unfamiliar with Tiny Corp, the important thing to know is that they're responsible for building "TinyBox", which is their design for an AI-powered server using AMD hardware. While the TinyBox indeed shows the potential power of AMD's raw GPU compute, deeper issues seemed to force Tiny Corp to make an Nvidia version, as well, citing that the AMD machine simply doesn't work as intended.

The ongoing updates to AMD's ROCm platform alongside this news should hopefully mean that the AMD TinyBox is competitive as an AI solution against Nvidia sooner rather than later. Since MES was the most important thing that needed to be open-sourced according to TinyCorp, the resultant bug fixes may be just enough to make the AMD TinyBox (and AMD AI hardware in general) look more competitive against Team Green. As it stands, it seems like nearly everybody is hoping to undercut Nvidia's dominant position in AI hardware and pro workloads.

According to the original AMD Twitter post, the MES documentation and source code release is slated for the end of May, and will include a GitHub tracker for easy collaboration with the community. 

At the time of this writing, TinyCorp and George Hotz have yet to make any kind of official comment on this announcement from AMD, but most recently highlighted their "TinyBox green" and "TinyBox red" kits for pre-order, emphasizing that the AMD box achieves the same compute for $10K cheaper but expertise will be needed to get it working. The phrasing used is "Make AMD great, with or without their help"— fortunately, it seems like AMD will be helping after all.

Freelance News Writer
  • The_Future_Is_Here
    I feel like the slow reaction from AMD actually responding was probably due to AMD having to ensure it met all legal obligations to both it's partners and it's investors, a lot of red tape there to ensure nobody tries any funny business & gets away with it (although it does still happen, don't doubt that, just maybe not at AMD).

    It is clear that AMD does indeed have the hardware designs needed to compete with Nvidia, but quite clearly (and this has been clear for a few years now), AMD seriously lack the necessary software expertise & software developer investment to give itself an opportunity to truly compete against the Goliath that is Nvidia in the current AI development & AI Workloads space in the market.

    Ultimately the winners of this newly opened up market will be the ones who are most Open with their software & hardware stacks, and the players that offer the easiest to understand documentation & software solutions that make deployment as easy as possible.
    If any of the big players don't have both of those cards in their hand then they will lose out in a big way as the market grows & expands to encompass different sub-markets and new opportunities not yet considered.

    Nvidia right now has the proprietary software advantage, but for how long they can continue to hold onto the market with that advantage we don't yet know, if any other players enter the ring with similar performance options for a similar price but with an Open-Source architecture platform then Nvidia may see themsekves being dethroned within just a couple of years.

    The "AI" (Hardware Accelerated Low-Precision Algorithmic Workflows) market is rapidly evolving and to be quite honest it's still really anyone's market to dominate or exploit, theirs plenty of opportunity and a lot of the already existing players have made it clear that they are incredibly slow to adapting & moving in the same direction as a market that's able to change this fast.

    A well placed, well timed, &, well executed market infiltration from a smaller player like TensTorrent for example could easily unsettle the Goliaths and make for a really quite exciting market for all.
    Reply
  • jeffy9987
    i like opensource like the next guy but doing this so soon to tinycorps whining really makes amd look weak i wished they would have waited another month or two
    Reply
  • vijosef
    Nvidia has an advantage that nobody else can match, except AMD (who wastes his opportunity).

    Nvidia's advantage is that teenagers who own a GPU for gaming, youtube streaming and other consumer tasks, can start doing serious work with it, and if they become AI developers, they enter the working force with a decade or more of experience in nvidia hardware and ecosystem.

    Instead any other competitor needs to start training users from zero at the day they start working (unless they positioned their hardware at some university), since most of their hardware needs servers and/or are only available for enterprise.

    So, even if somebody makes better hardware than nvidia, the programmers do not have the experience to squeeze it.
    Reply
  • bolweval
    jeffy9987 said:
    i like opensource like the next guy but doing this so soon to tinycorps whining really makes amd look weak i wished they would have waited another month or two
    This isn’t War & Peace, it’s AMD helping out a customer solve it’s own problems, ain’t about looking weak…
    Reply
  • bit_user
    In the meantime, Tiny has already decided to bypass MES. So, this move won't help Tiny or probably even AMD.

    Most likely, it'll just provide valuable insight for certain upstarts trying to build their own GPUs about how AMD does thread scheduling and other GPU internals.

    This whole affair has been nothing but bad for AMD. If they'd politely rebuffed Hotz' overtures, instead of letting him string them along, they'd have probably even had less negative PR from it.
    Reply
  • NinoPino
    bit_user said:
    In the meantime, Tiny has already decided to bypass MES. So, this move won't help Tiny or probably even AMD.

    Most likely, it'll just provide valuable insight for certain upstarts trying to build their own GPUs about how AMD does thread scheduling and other GPU internals.
    I doubt AMD would have done so if something really significant would be disclosed.

    bit_user said:
    This whole affair has been nothing but bad for AMD. If they'd politely rebuffed Hotz' overtures, instead of letting him string them along, they'd have probably even had less negative PR from it.
    Maybe the AMD move is both to help dev community and at the same time remove the Hotz alibi.
    Reply