Tiny Corp At "70%" Confidence For AMD To Open-Source Some Relevant GPU Firmware

[coder]

That's true, but is this actually cut out from the chip, or just locked? I was under impression, that they just lock it, since making different chips is costly.

No, the consumer chips don't have it. Study this page:

https://en.wikipedia.org/wiki/List_o...ocessing_units

It shows you what die each GPU is made from. You can see that after the GK110 (GTX 780 Ti, Titan, Tinan Black, Titan Z), their fp64-heavy dies were never again used in any gaming cards (with the exception of the Titan V, which I don't consider a gaming card, due to its $3k price tag).

AMD, on the other hand, can leave it at 1:2, as there is no software to run on these cards anyway.

AMD also doesn't want to waste die area on an unnecessary feature. The last AMD gaming card to have it was the Radeon VII, which they did nerf down to 1:4, whereas the hardware was actually 1:2. Keep in mind that AMD said the Vega 20 chip wasn't designed as a gaming GPU and only shipped as one because they saw a market opportunity for it.

After that, AMD split their HPC and client GPUs into two different architectural families. CDNA is a very different architecture than RDNA. Only the CDNA dies have vector fp64. The RDNA dies just have scalar fp64 (like Nvidia's client GPUs), but they now have a 32-element wavefront size. This explains their 1:32 ratio.

Save for OpenCL which is nowhere as popular as it should be.

100% agree.

I wish they opened this ecosystem to the point community could provide fixes, optimizations, and bring it to any Radeon chip it can run on. Just like let's say RADV. But I don't set my hopes too high. Maybe Intel will push them a bit.

I think I read that some crypto miners were using custom firmware. I wonder if it was reverse-engineered, leaked, or maybe they even managed to get the necessary tools & docs through a partnership agreement with AMD.

[bridgman]

I think I read that some crypto miners were using custom firmware. I wonder if it was reverse-engineered, leaked, or maybe they even managed to get the necessary tools & docs through a partnership agreement with AMD.

I believe we supplied the firmware, although IIRC it was primarily power management tweaks.

[danilw]

Interesting to read.

Especially that Nvidia now thinks "consumer should not run AI on their consumer GPUs" making limitations to CUDA.
Will be interesting to see how Nvidia will require monthly subscribtion to unlock CUDA on consumer GPUs.
While AMD opensource - "unexpected two trillion dollars corporation flop".