AMD Says They'll Be Open-Sourcing More Of Their GPU Software Stack & Hardware Docs

[fallingcats]

AMD is probably spending too much time working with Valve these days.

I doubt they're spending all that much time on Valve, except maybe fixing bugs that affect them with a higher priority. In the grand scheme of things, valve isn't selling a whole lot of devices compared to AMD, which are in almost all major consoles.

[pieman]

AMD: We are open sourcing more of our code!

FOSS Hypocrites : Not open enough! ....typed from their Ngreedia equipped computers...

• CPU: 13th Gen Intel i9-13900K (32) @ 5.5GHz
• GPU: AMD ATI Radeon RX 7900 XT/7900 XTX

typing this from my 7900 xtx. i love amd gpu's. i just want them to have MUCH better driver reset / recover. its really my only true complaint at this stage. one bad driver crash and reset leaves my computer in a lock state... its not fun. i don't know how a reset still lands in a frozen system.. luckily i haven't had a crash in awhile. the only game causing it was world of warcraft (not really a linux issue, windows has the same exact issue) my "fix" was not using radv and using amdvlk just for wow only. got the idea after seeing windows users mentioning using dxvk on windows fixed their wow crashing radeon drivers so i figured it had to be their vulkan driver that was preventing the crashing. and yeah, there is an open issue on mesa about it. that rant aside, everything else has been good to me. i love my 7900 xtx.

[ahrs]

Nice, glad to see the tinycorp stuff at least made some progress, Lets see if their compute situation will actually improve or if this is just fluff

If this results in anything other than the firmware being open sourced, it's just fluff. Tinycorp already replaced all of AMD's userspace. There is no "AMD software stack" in their repertoire.

[pong]

I'm glad (wrt. AMD, Intel, and to the extent it applies, NVIDIA) things are as open as they are.
I'm further glad things work at least as well as they do so one can get a baseline level (more or less) of compute / graphics support from GPUs from each of the designers.

But it's so strange on the one hand choosing to "support open source", "support linux" and yet in every case AFAIK there are glaring gaping holes that would be low hanging fruit that's easy to resolve and yet it just sits unaddressed for long periods.

For instance AFAIK there's no LINUX sensor (temperature, voltage, current, power, clock, fan speed) support uniformly for each vendor in terms of a CLI utility, GUI utility, underlying APIs, and relevant HW documentation. I think nvidia had the better settings utility / UI story (AFAIK) wrt. nv-smi, nvidia-settings, though the latter (AFAIK) doesn't work at all without X and that's pretty crippling. I don't personally know of AMD / Intel made LINUX CLI / GUI utilities for those things. And it's not clear to me the low level HW documentation / driver-control layer APIs are published either.

Though since there are some 3rd party FOSS clock / fan / temperature / power related UIs for AMD maybe they've documented something either at the API or the HW level these things are using unless it's all independently analyzed & implemented.

And they all incredibly (and hypocritically with much cognitive dissonance) support "virtualization" and "isolation" with MMU / IOMMU / VM etc. in their CPUs and yet utterly reverse course and go out of their way to break standards-based SR-IOV functionality on the GPUs which would allow better virtualization (at least for a few VMs / isolated compute contexts) on consumer models.

I think the "spirit" of FOSS isn't so much about free-as-in-beer, nor even software, underlying I think it's about having IT tools that have known architectures, capabilities, extensibility (bring your own code / patches / whatever), composability (SW, open standard hardware interfaces), and ability to maintain / make full use of the SW / HW you've invested in to the maximum extent possible even if it means you're the one doing the work of maintenance / extension vs. the HW vendor.

Now we've got phones / tablets / smart watches / IOT stuff so locked down you're not even "root" of the HW you own, can't write / install your own SW on it in many cases. We've got increasingly "black box appliance" computing (apple) systems where you can't use key parts of its capabilities or program / use it according to your own wishes however unreasonable those restrictions are against your own customers.

And IMO the undiscussed elephant in the room is still that we're all salivating for better GPU capabilities when the primary / only reason for wanting those are
really to have a parallel computing capability that has modern memory BW O(1TB/s) and modern SIMD parallel capability O(8k "cores") and the very
chip makers who are (or will be) controlling / manufacturing our core computer architectures (intel, amd, nvidia cpus / chipsets) are sort of like a cartel keeping
the PC / ATX platform stagnated so today's "enthusiast desktop" still has less 2-channel DRAM BW than a ~20 year old consumer GPU (e.g. 8800GTX) had VRAM BW
and we've got nothing like GPU SIMD processors as core or holistically extensible computing elements on the PC due primarily to the ongoing market
segmentation of AMD, Intel, and obviously Microsoft's continued push to dumb down the PC and monetize the users vs. advance the SOTA for the platform.

[BaumKuchen]

DItched Nvidia/intel year ago and never looked back so far.

[Quackdoc]

If this results in anything other than the firmware being open sourced, it's just fluff. Tinycorp already replaced all of AMD's userspace. There is no "AMD software stack" in their repertoire.

is there any evidence for this? afaik tinygrad is using rocm

[gentoofu]

is there any evidence for this? afaik tinygrad is using rocm

https://twitter.com/i/web/status/1764737741693342128

What does the Hotz bath water taste like?

[Eirikr1848]

AMD is probably spending too much time working with Valve these days.

AMD hardware is in the PS4/PS5 and Xbox One/One S/One X; Series S/X.

Undoubtedly the Steam Deck takes up some resources, but it is quite imaginable other projects take up more

[Quackdoc]

https://twitter.com/i/web/status/1764737741693342128

What does the Hotz bath water taste like?

Oh that's neat, I wonder how this is getting implemented in tinygrad, I am now seeing the comgr commits, but I thought they were just part of the rocm optimization, also seeing some commits with "KFD" I wonder if that is related.

impression was due to the writeup for https://github.com/tinygrad/tinygrad/pull/3851

EDIT: though the comgr stuff is userland and open source, so it's not wholly true that all of the userspace stuff have been replaced, unless comgr is also being replaced. https://github.com/ROCm/llvm-project...ging/amd/comgr

[Anux]

So a tiny corporation saying their closed firmware is shit is enough for AMD to open source it? All those years the community begged for it and they were just ignored.