AMD Clarifies ROCm Compute Support For GUI Applications

[Guest]

ROCm is a joke, I hope clover gets some traction.

Or clvk instead.

[_ONH_]

... , I hope clover gets some traction.

How could it the stats page of it is frome sometimes in 2018.
so he package thrown over the fence is the best sollution available.

[darkbasic]

How could it the stats page of it is frome sometimes in 2018.

Because of this: https://gitlab.freedesktop.org/mesa/..._requests/7241

[ms178]

linux is far more important than the people who are used to prever windows/mictosoft would admit.

I doubt that very much, at least today, it also depends on the market. The desktop Linux market is still trailing Windows by far, that's where most gaming cards are targeted at. But the importance of the HPC/enterprise is expected to grow significantly over the coming years were Linux is way ahead, today GPU sales are still behind CPU sales in that sector but the proportions are about to change (source with analysis).

Especially that part from that article says it all: "Right now, we reckon that AMD is getting two thirds of its datacenter revenues from Epycs and one third from Radeon Instincts. But in the fullness of time, the revenues should reflect the ratio of CPU to GPU compute normalized to prices, and that should mean a lot more revenue coming from GPUs than CPUs. So by the time we get to 2023, the revenues from GPUs could be a lot larger portion of the AMD datacenter pie. Given the 1:4 ratio of the exascale-class supercomputers, “Frontier” and “El Capitan,” being built by Cray for the DOE using AMD Epyc and Radeon Instinct motors, and assuming that GPUs will cost more than CPUs in these machines, the fact that AMD is winning the GPU portions of these hybrid machines is truly significant for its top and bottom line in its datacenter business."

[ms178]

i said all perfectly about HPC/enterprise.

You were very broad and generic in you choice of words, but I think we agree on that their HPC/datacenter part is getting more important for them in terms of revenue and therefore the importance of a good Linux stack for them - maybe I should have mentioned it.

means many developers who are in the "HPC/enterprise" sector also use linux on their desktop.
this means linux desktop is not some independent thing instead the markets have interconnections.

Can you back that up with any data? I mean it is safe to assume that super computers and enterprise servers are not used by people directly for gaming/desktop usages at all. So this leaves you with workstation usage scenarios on Linux, I doubt the relevance in terms of overall revenue in that niche for AMD.

also gaming/desktop market in the last steam statistic intel lost 2% marketshare to amd on linux.
but on windows this marketshare shift to amd did not happen.
i am 100% sure more and more linux desktop users use AMD hardware.

for example on windows the most used gpu is "NVIDIA GeForce GTX 1060" with
9.52% on linux instead ot is the AMD "
AMD Radeon RX 480
7.58%"
this means AMD won the linux desktop market already... on windows they have no chance at all and on linux they win.

I wouldn't read too much into these Steam numbers to begin with, it all comes down to the question of how truely representative they are. If your Chinese gaming cafes are counted equally as normal desktop users, that would be skewing the results in one direction already. Also the driver string reporting seems to be different on Windows and Linux (as the RX480 on Windows is likely to be reported as RX580 - at least it is odd that the RX580 is far more popular on Windows than Linux). You should notice that the sample size is small which means that small differences in absolute numbers have a higher impact in that group. The shift of Intel to AMD in CPUs you talk about might not be significant at all when considering the bigger overall picture.

AMD might have "won" the linux desktop market, but do not get carried away from the percentages here again as the overall absolute numbers are still tiny in comparison to the Windows market. As GTX 1060 and RX 480/580 were competing with each other and had similar performance this should lead to roughly equal market share if price and availability are assumed to be also in the same ballpark. As you see from the numbers the RX480 has a slight lead over the GTX 1060 on Linux which equals to 1,4:1 in favor of the AMD card wheras on Windows it is not even close, 4,53:1 for the Nvidia card. As the overall market share of Windows is the dominant factor, the latter numbers are way more important for the companies and their sales figures.

Also it is not guaranteed that the positive trend for AMD on Linux would continue if more and more users would switch over to Linux - this is what you wanted to imply, wasn't it?! If more Windows users decided today to make the switch to Linux without changing their hardware, you simply would see a decline in RX 480 numbers and a rise of GTX 1060 numbers.

[Spacefish]

Well, one motivation to have good Compute Stack support on consumer hardware under linux: Academia / new programmers.
Typically people working in the datascience / machine learning field work with linux / ubuntu. Running most machine learning frameworks / tools on Windows is somehow "unsupported" / "yes we have win32 support, but please go away with your questions, no one uses it like that".

99% of people doing udacity courses / learning these things, start on their local PC. Typically you don´t buy a MI100 just to try it out, especially not as a student..
That´s why a lot of people buy a nvidia GPU if they want to get into machine learning and run some smaller tasks on their PC..

Supporting consumer SKUs in ROCm would be an investment into the future / into the ecosystem for AMD.
I don´t think the HEDT market for Linux systems for creative work (3D Rendering / Video Editing) is especially developed ^^. These "creative" guys typically use that Darwin based OS and Windows...

As far as i can tell from the bugtracker, ROCm support for RDNA is coming this year! So let´s hope this holds true!
Once RDNA/RDNA2 works for machine learning tasks, i am sure we will see a lot of kernel optimization merge requests and so on, as soon as people start to use it!

Furthermore, i still think it will drive sales for HPC in the future, once people getting into the market today / learning things today start to work in large enterprises.

[ms178]

this means it is more important what kind of customers you have compared to how many customers you have.

sure if you only count market share you go for windows but if you count what kind of customers you have then you find a magic bulled or the Holy Grail means you find out that your linux desktop/gaming customers are the same people who also work in professional computer science field...

this means the linux desktop market is not disconnected and indepentend instead it is connected to the HPC server market...
because it is the same high value customers... the only difference is that the HPC server run at "work" and the steam gaming desktop run at "home"... but it is the same people.

Thanks for clarifying that point, I now get what you and Spacefish wanted to say. I agree partly, while this might be important for AMD's mindshare with academics and their impact on HPC/enterprise down the road, at least for now for every one of them there are dozens of casual gamers buying AMD products, that still makes Windows a priority but I agree that good Linux support has become more important over the years as they want to compete with Nvidia (and soon Intel) in that segment. But I'd say good products and delivering on roadmaps is what contributed largely to their HPC design wins as their software ecosystem still lacked behind Nvidia's when these buying decisions were made, hence the software side might not have been the deciding factor (or at least ROCm and the rest of their stack seemed to be good enough to build or improve upon).

Spacefish, I agree with that statement in general, I also fear that it is not a good sign that ROCm lacks behind in support for RDNA. Maybe it was just the amount of work needed, but appearently it wasn't a bigger priority to make it happen sooner. I don't get tired to admit that I am more of a fan of their Vega approach which was a good compromise of both graphics and compute capabilities in one product, but maybe they will offer us consumers the best of both worlds again when chiplet-based CPU/GPU/accelerator combos become a thing in the future. Until then, it seems the split of their products will be more severe and you need to buy one of their professional product for the best hardware and software support for these tasks.

[Spacefish]

ms178 from a technical standpoint it was the right thing to do to split these products, as there are different requirements..
Especially with the Raytracing functionality, you "waste" a lot of chip area you don´t use in HPC typically. Furthermore HPC applications don´t need all the display related hardware, video / image encoders / decoders and so on.. Spend that chip area on wider buses more compute units and the money on memory .
From a marketing / ecosystem standpoint it was a bad decision IMHO.

[bridgman]

I agree with that statement in general, I also fear that it is not a good sign that ROCm lacks behind in support for RDNA. Maybe it was just the amount of work needed, but appearently it wasn't a bigger priority to make it happen sooner.

The key question is pretty simple - what should the MLSE group have *not* done in order to free up budget for RDNA ROCm support. ?

By the end of 2020 we had just finished building up the ROCM stack to the point where it was sufficient for most prospective customers - we would have had to delay that for a year or so if we prioritized RDNA support, which in turn would probably have cost us a number of business opportunities. Do you think that would have been a good decision ?

For reference see bottom of page 7 of CDNA white paper: https://www.amd.com/system/files/doc...whitepaper.pdf

If you are thinking "well it would have been better if you had had enough money a couple of years ago to be able to do both" I'm sure we would all agree with you.

There were some compromise things we could do, like bringing up the bottom of the ROCm stack on all our new hardware to make sure that any HW issues were found early and to lay the foundation for moving OpenCL from PAL to HIP a couple of months ago - and we did those - but for the upper layer components it was more of an either-or unfortunately.

[Spacefish]

bridgman thanks for the insights!
Hope the recent developments increased the budget a little bit and the money currently earned is reinvested into R&D.
Guess it´s always a balance between future looking investments and investments that earn you money right now / in the short term.