Lisa Su Says The "Team Is On It" After Tweet About Open-Source AMD GPU Firmware

[phoronix]

Phoronix: Lisa Su Says The "Team Is On It" After Tweet About Open-Source AMD GPU Firmware

George Hotz with Tiny Corp that is working on Tinygrad and TinyBox for interesting developments in the open-source AI space has previously called out AMD over ROCm issues. Yesterday yielded new tweets by "the tiny corp" over AI training runs crashing with MES errors and then called for AMD open-sourcing the firmware to which AMD CEO Lisa Su has responded...

https://www.phoronix.com/news/Lisa-Su-Tweet-OSS-Firmware

[Kjell]

First open source BIOS and now GPU Firmware..

TAKE MY WALLET

[Quackdoc]

I trust AMD for compute as much as I would trust charlie scheen's used condom to be sterile, needless to say, none of the customers i've ever had, or ever will have, have used, or will use AMD in the near future.

[oguzoztoprak]

I trust AMD for compute as much as I would trust charlie scheen's used condom to be sterile, needless to say, none of the customers i've ever had, or ever will have, have used, or will use AMD in the near future.

That is a bold statement. I hope AMD takes a turn for the better. We need more competition.

[Quackdoc]

That is a bold statement. I hope AMD takes a turn for the better. We need more competition.

Im actually pretty hopeful for Intel. I use VS a decent amount and recently bm3dcuda got an intel sycl implementation and it works fairly well. I tried to use the rocm version but only 1/5 people who I had test it got it working xD. Needless to say I dont have much hopes from AMD, Intel on the otherhand seems actually fairly promissing.

[rommyappus]

First open source BIOS and now GPU Firmware..

TAKE MY WALLET

I don't think she means to open source the firmware but to deal with the firmware issues.. so i guess we'll see.

[Jumbotron]

The tweet that just might change things. It brings back into the spotlight what Lisa Su stated a year or so ago about "redoubling efforts" ....(my paraphrase)....concerning ROCm. I took what Lisa said as ...."I understand that ROCm is crap after all these years after scrapping HSA and the work done with Fusion architecture before I joined but we needed to make a great server CPU again to get our profit margins up in order to compete with Intel in the server and HPC space."

Say what you will about Nvidia...they have their software stack tighter than Dick's Hatband and even though closed source they were smart enough to get CUDA into academia so a whole generation or more of developers know nothing else....or nothing else quite like CUDA as opposed to dinking around with OpenCL and now ROCm..

I am REALLY...REALLY hoping that with the introduction of Zen 6 CPU arch and RDNA/CDNA 5 GPU arch along with the latest Infinity arch and fabric as well as a finally fleshed out and easy to install and manage ROCm, not to mention AMD's acquisition of Xilinx, that the promise of what the Fusion arch and HSA vision that AMD started way back in 2011 finally comes to pass. We had that for a brief shining moment with Kaveri and later optimized with Carrizo and finally Bristol Ridge. I still have two computers with Bristol Ridge APU's. The Playstation 4 had that. Zero Copy of memory between CPU and GPU. Unified 48 bit memory addressing between CPU and GPU. Everything connects to everything. At least on paper. It's just a shame that what AMD created for the first time in the x86 world, it took Apple to perfect and release to the public in the form of their Apple Silicon chips...at least for desktop purposes.

At this point, if Lisa can't get her teams to make extraordinary advances in ROCm and clean up the mess that is ROCm then she should just abandoned the whole project and wholeheartily adopt and extend Intel's oneAPI and Intel's entire compute stack and call it a day.

[pWe00Iri3e7Z9lHOX2Qx]

I don't think she means to open source the firmware but to deal with the firmware issues.. so i guess we'll see.

Yeah, this happens all the time. Bad press. Big boss tells one of their directs to fix the specific issue in the bad press. Rolls down hill. Some engineer who is probably on PTO somewhere ends up spending half his/her vacation fixing the "critical" issue since they know that code the best.

[ssokolow]

I suppose an intent to open-source the firmware would justify not doing HDMI 2.1 in the firmware blob.

[pWe00Iri3e7Z9lHOX2Qx]

It's still hard to take AMD seriously on any comments related to GPU compute. There was that 'big things are coming soon for ROCm we pinky swear!' message. George's message is accurate but applies at a much higher level. 'FIX YOUR SHIT! ' applies to their entire strategy around GPU compute.

AMD market cap: $340B
NVIDIA market cap: $2.2T

Basic logic says that as the x86 CPU leader (at least in terms of current tech/arch/performance/efficiency) AND one of the two big GPU vendors, those numbers above should be reversed. And that $2.2T market cap has basically nothing to do with desktop or gaming. GPU compute rules the day. AMD's constant fail here is hard to understand. NVIDIA already wrote the playbook. The ROI is obvious. So indeed, AMD, fix your shit.