AMD ROCm 6.1.1 Brings Fixes, Preps For Upcoming Changes & cuDNN 9.0 Support

[pipe13]

Isn't it obvious? Money and manpower. People buy Intel no matter what they do. Thus, they can pay more people to work on the GPU stack.



By the way, ROCm works fine on my consumer hardware as well.

Er, and just what hardware do you consume?

[oleid]

Er, and just what hardware do you consume?

I don't consume any hardware, as I'm particular when it comes to my diet.

[aerospace]

gfx1030 at least has inofficial support. Navi1 (RX 5700) is gfx1010 and not supported at all. It used to work with the gfx1030 workaround, but that doesn't work (for me) since a long time. AFAIK there now is generic code that should support 1010 as well, but it isn't being shipped with ROCm 6.1.0. I couldn't test 6.1.1 yet, but my hopes are low.

Are you aware of 780M (should be gfx1103) support?

[tenchrio]

gfx1030 at least has inofficial support. Navi1 (RX 5700) is gfx1010 and not supported at all. It used to work with the gfx1030 workaround, but that doesn't work (for me) since a long time. AFAIK there now is generic code that should support 1010 as well, but it isn't being shipped with ROCm 6.1.0. I couldn't test 6.1.1 yet, but my hopes are low.

Looked it up, I remember I checked for someone else a while back and how GFX1010 was being blocked due to some kind of bug/incompatibility that AMD was still working on. Should be implemented by ROCM 6.2, as stated later down Debian had already enabled GFX1010 in their own build. Honestly the Debian docs give you a better list of supported GPUs/HSA, with the exception of GFX101*.

Only downside to the Debian builds is that the Debian-AI team is still working on making their own Pytorch and MIOpen builds so you would have to build these yourself. As an alternative you could use the rocm-xtra docker builds (which also offers pre-configured docker builds for Pytorch, Stablediffusion and LLAMA).
For anyone interested, it seems rocm-xtra also brings compatibility back to GFX803 (so RX580 most likely you can also run other Polaris cards by targeting this).

Is RX 6800 base supported (gfx1030) or not?

It is. The HSA is shared with the AMD Radeon PRO W6800, because of this every GPU with that HSA/LLVM target is usable with ROCM.
Why AMD doesn't outright list the RX 6800 series has never really been clarified, my guess would be for legal reasons.

[pWe00Iri3e7Z9lHOX2Qx]

Isn't it obvious? Money and manpower. People buy Intel no matter what they do. Thus, they can pay more people to work on the GPU stack.

Except this isn't a remotely valid excuse anymore. 8 years ago when ROCm was first release you could make this excuse. Not now. People need to get over thinking that AMD is broke.

Intel market cap: $128B
AMD market cap: $247B

Intel debt: $49B
AMD debt: $3B

Intel cash on hand: $25B
AMD cash on hand: $5B

AMD's financials are better than Intel's.

[cgmb]

No consumer hardware GPU support yet?

The RX 7900 XTX, RX 7900 XT, and RX 7900 GRE are on AMD's official support list, and in practice the Radeon VII, RX 6800, RX 6800 XT, RX 6900 XT, and RX 6950 XT work fine too. As well, some distros have enabled built-in support for more hardware. For example, the distro-provided packages in Debian Unstable and Ubuntu 24.04 are enabled on all discrete Vega, RDNA 1, CDNA 1, RDNA 2, CDNA 2, and RDNA 3 GPUs.

[tenchrio]

Except this isn't a remotely valid excuse anymore. 8 years ago when ROCm was first release you could make this excuse. Not now. People need to get over thinking that AMD is broke.

Intel market cap: $128B
AMD market cap: $247B

Intel debt: $49B
AMD debt: $3B

Intel cash on hand: $25B
AMD cash on hand: $5B

AMD's financials are better than Intel's.

People need to stop using market cap, it has nothing to do with a companies actual financials.
In 2023 Intel still had the largest Revenue Stream of the two ($54.2B), AMD only had less than half of that ($22.68B).
More in depth analysis of their financial reports reveals that this also follows in suite with their R&D spending with Intel ($16B) spending a lot more than AMD ($5,8B) as they can due to a larger revenue stream (and a lot more asset value as can be seen in both financial reports). So no, Intel still has better financials over AMD.

[HighValueWarrior]

I love me some AMD .... but the apologists are getting out of hand.
Get your shit together AMD devs ... it is embarassing now.

[pWe00Iri3e7Z9lHOX2Qx]

People need to stop using market cap, it has nothing to do with a companies actual financials.
In 2023 Intel still had the largest Revenue Stream of the two ($54.2B), AMD only had less than half of that ($22.68B).
More in depth analysis of their financial reports reveals that this also follows in suite with their R&D spending with Intel ($16B) spending a lot more than AMD ($5,8B) as they can due to a larger revenue stream (and a lot more asset value as can be seen in both financial reports). So no, Intel still has better financials over AMD.

Nobody was just looking at market cap. Intel has a large amount of debt that is accelerating. Their earnings before interest and tax dropped 99% over the last year. They are not in great shape. If being first to backside power delivery with Arrow Lake doesn't make them much more competitive with AMD, I'm not enthusiastic about the future. They need some big foundry service wins.

[qarium]

Nobody was just looking at market cap. Intel has a large amount of debt that is accelerating. Their earnings before interest and tax dropped 99% over the last year. They are not in great shape. If being first to backside power delivery with Arrow Lake doesn't make them much more competitive with AMD, I'm not enthusiastic about the future. They need some big foundry service wins.

backside power delivery does not make the transistors faster or more efficient.
it has the same effect as intels 10nm already has intels 10nm already has the transistor density similar to 7nm TSMC...
but no one consider intels 10nm be on par with TSMC 7nm even if intel has the transistor density...

same with backside power delivery it increase the transistor density does not make the transistor faster or more efficient.

also keep in mind that they really need this because High NA EUV is much more expensive than EUV...

this means without higher transistor density High NA EUV could result in a cost explosion.

keep in mind at TSMC even 1,6nm or 1,4nm node will not have High NA EUV because of the cost explosion.

so your hope: "make them much more competitive with AMD" is clearly not happening their tranistors will still be slower and less efficient in power use.

all what backside power delivery does is bring down the cost of High NA EUV by increase the transistor density..