Lisa Su Reaffirms Commitment To Improving AMD ROCm Support, Engaging The Community

[Developer12]

I don't quite understand what's the main problem with ROCm. Is it a complicated install/deploy procedure or is it being virtually non-functional?

The former one is a semi-issue i'd say. To this date AMD is more of a hardware/components company, not a complete solution company. IMHO it's OK to provide only core tools/framework support instead complete solution if you price your hardware accordingly. It becomes a problem only when AMD tries to position itself as a equivalent competitive premium brand but actually lacks competitiveness in a software/features.

it works on like 5 cards and none of them are consumer cards. you can try to run it on like 5 more consumer cards but it's a crapshoot whether it'll actually work.

[smitty3268]

From what I can tell for the losses, it's a mix of processor sales being down due to AM4>AM5 combined with right now being the traditional lull between game console generations.

I'm not sure if this is part of the P&L statements, or even if it actually happened, but I believe AMD committed to some giant ($2 billion?) stock buyback program this year.

So while it might be fair to point out that they aren't raking in unlimited amounts of cash, I think it's also fair to suggest they have plenty of money available to fund drivers if they believe it's important to the company.

Edit - it was $8 billion, with no exact timeline on when purchases were happening. Just 1% of that is 80 million, which could fund quite a bit of driver work.

[abott]

Even if they don't make 2 billion in profit a quarter, they obviously can't afford to have their software suck as much as it is, point blank.

[cgmb]

it works on like 5 cards and none of them are consumer cards. you can try to run it on like 5 more consumer cards but it's a crapshoot whether it'll actually work.

I know that's what it feels like, but that's more a reflection of ROCm's limited documentation on hardware compatibility than the actual truth. ROCm only officially supports a handful of server or workstation cards, but "officially supports" is not the same thing as "works on". I haven't finished testing all AMD GPUs, but thus far it seems to me that HIP works on all GFX9, GFX10 and GFX11 GPUs. With the right magic incantations, you can get the math and AI libraries working on most of them, too. My experience has been that ROCm will work on almost all AMD hardware from Vega onwards, if you can figure out the right settings.

Most libraries in the AMD-provided binary packages run on gfx900, gfx906, gfx908, gfx90a, gfx1030, gfx1100, gfx1101 and gfx1102. With the right environment variables, the libraries will also run on gfx803, gfx902, gfx909, gfx90c, gfx1031, gfx1032, gfx1033, gfx1034, gfx1035, and gfx1036. As far as I can tell, the only GFX9, GFX10 or GFX11 AMD GPUs that you can't get working with all the prebuilt libraries just by setting some environment variables are the RDNA 1 GPUs and the RX 7600 (in ROCm 5.5 at least).

I've been packaging ROCm for Debian and I've documented the relevant environment variables in the Debian Supported GPUs wiki in the Architecture Notes section. It's was a pretty hastily-written document, but I hope folks find it helpful.

[CochainComplex]

I run a decent HPC at well known institute so I know how these things work behind the scenes. Right now it is not possible to get consumer cards into server chassis anymore due to thermal constraints, nvidia made sure that's the case. And whole HPC world is panicking as nvidia datacenter cards kept perf/$ from gen to gen and lots of places simply cannot afford them anymore. Also nvidia is happy selling their cards to industry and doesn't waste time on scientists begging for academic pricing etc...
So market is ripe for a disruption and if AMD is smart, they will attack full force. Lots of these well knows institutes already have some Mi210 cards here and there and are porting their stuff and I expect Mi300 to sell very well as it will bring the next step in perf/$ and perf/W. (Nobody cares about absolute performance of single gpus anymore)

Thx for the insight - I#m only frequently the user of a HPC so I don't know anything about the commisioning. I should have known better that Nvidia is playing the same game in HPC as they do with gamers.

[vein]

This forum is so funny... I always need to bring some popcorn when you read it.

I have been running ROCM on my fury since the..beginning..it was some problems in the beginning some since a while it has been in official repos in arch and there is now problems on my old fury or my newer cards.. Everything just works in blender or pyTorch or whatever..

But I guess haters gona hate...

[Grinness]

This forum is so funny... I always need to bring some popcorn when you read it.

I have been running ROCM on my fury since the..beginning..it was some problems in the beginning some since a while it has been in official repos in arch and there is now problems on my old fury or my newer cards.. Everything just works in blender or pyTorch or whatever..

But I guess haters gona hate...

Same here, ROCm works on rx6800 (rdna2) and rx 480 (polaris), blender, pytorch and so on -- on Arch with official packages
Yes it does not work on APUs (at the moment I would add) - but this is a red herring: nobody mention that you need GPU RAM to use any decent compute (AI) app (8GB+ I would add)

This forum is absurd.
Note that I do welcome AMD efforts in standardizing and easing the installation process and expanding official support.

[skeevy420]

I'm not sure if this is part of the P&L statements, or even if it actually happened, but I believe AMD committed to some giant ($2 billion?) stock buyback program this year.

So while it might be fair to point out that they aren't raking in unlimited amounts of cash, I think it's also fair to suggest they have plenty of money available to fund drivers if they believe it's important to the company.

Edit - it was $8 billion, with no exact timeline on when purchases were happening. Just 1% of that is 80 million, which could fund quite a bit of driver work.

I was just going by their percent of change. Consumer sales and consoles were down.

AMD obviously has cash on hand. Just because they had one slightly bad year doesn't mean they didn't rake in money during the pandemic when prices skyrocketed and they couldn't keep up with demand or the few years prior when they took the world by storm with Ryzen which increased their stock prices more than 10 fold.

They also spent a ton of money, $49 Billion, acquiring Xilinx in 2022....which was their most profitable division at $1.3B in 2023. $8 Billion on some Ronald Regan Era Trickle Down Stock Buybacks using money earned from Trump Era Corporate Tax Cuts is a drop in the bucket compared to what was spent on Xilinx.

As an AMD user, I'm a bit upset that they can basically spend $57 BILLION DOLLARS on themselves and, yet, they can't seem to find any money to hire a person to make a damn GUI or a universal Pro driver (not that I'd use Pro, Mesa/RADV FTW, but the sentiment counts).

[veritas]

I was just going by their percent of change. Consumer sales and consoles were down.

AMD obviously has cash on hand. Just because they had one slightly bad year doesn't mean they didn't rake in money during the pandemic when prices skyrocketed and they couldn't keep up with demand or the few years prior when they took the world by storm with Ryzen which increased their stock prices more than 10 fold.

They also spent a ton of money, $49 Billion, acquiring Xilinx in 2022....which was their most profitable division at $1.3B in 2023. $8 Billion on some Ronald Regan Era Trickle Down Stock Buybacks using money earned from Trump Era Corporate Tax Cuts is a drop in the bucket compared to what was spent on Xilinx.

As an AMD user, I'm a bit upset that they can basically spend $57 BILLION DOLLARS on themselves and, yet, they can't seem to find any money to hire a person to make a damn GUI or a universal Pro driver (not that I'd use Pro, Mesa/RADV FTW, but the sentiment counts).

Didn't AMD buy Xilinx using only stocks?

[bridgman]

Didn't AMD buy Xilinx using only stocks?

Yes, and the stock buybacks are in part to reduce the dilution (and stock devaluation) that results from acquiring other companies with stock.