AMD Quietly Funded A Drop-In CUDA Implementation Built On ROCm: It's Now Open-Source

[qarium]

Did you not read the comments to the phoronix link you cited? Btw, 8000 series are considered no better than mid-range cards. Perhaps, I should wait for RDNA 5?!? /s

well i remind you panix you told me you only want to spend 800-900€... and RDNA4 8000 midrange is not 7000 midrange.
we had this in the past that you claimed well a 7800 is to slow so then a 8800 is to slow to
but the card they do as 8800 is more like the 7900GRE means faster as the 7900GRE

and its easy to tell why a RDNA4 8800 will be better than a 3090 nvidia card... there will be a 32GB vram modell of the RDNA4 8800.-..
it is more or less a 7900GRE but they do the same as they did with the AMD PRO w7900 means they double the vram.

result is a 32GB vram RDNA4 8800 for like 800€

so from this perspective you will give yourself pain if you buy a 24GB vram 3090 instead.

[tenchrio]

Perhaps, you should debate ppl who know the material since I don't know anything. Even though, Puget only recommends - offers mostly Nvidia gpus in their custom builds for customers - you can check their site since you used them as a source, too. They also include a 4080 for AI work. There's this reddit page in which ppl debate using a 4080 vs a 7900 xtx - perhaps, you should lecture them, too? They probably know more than I do, though? If you read the thread, you'll know why I'm saying this. I just don't want to make the wrong decision when I buy a gpu.

https://www.reddit.com/r/StableDiffusion/comments/16nsdc0/4080_vs_7900_xtx_for_stable_diffusion/

I've even been advised to pick a (24gb) 3090 over a 7900 xtx - for AI/Stable Diffusion, Blender, Davinci Resolve (although, it's perhaps a bit more of a 'pick 'em') - it's only gaming in which the AMD gpu would be recommended. Why is that?

This generaton of gpus is honestly the most annoying when it comes to not making the wrong choice. Pascal (10 series) was definitely a massive go for Nvidia with a doubling in performance and VRAM compared to the previous gen, the GTX 1080 TI was such a winner of a GPU in price and performance that I would still confidently recommend it when Turing (20 series) released (especially since Optix was problematic at launch). With the 30 series it started to become harder, ray tracing and ai acceleration finally matured, I would argue that the RTX 3060 12GB is an easy buy as it was relatively cheap (especially after the crypto boom) and offered great performance, but the 3080 10GB felt ridiculous at the same price the GTX 1080 TI with 11GB once stood, the RTX 3080 Ti meanwhile felt overpriced at $1199 as the 12GB would clearly limit its performance (especially as 12GB felt entry level for 4K) and the 3090 was sitting at $1499 with double the VRAM meanwhile the performance difference for each tier of card shrunk compared to, oh say the GTX 1080 vs the GTX 1080TI (I cannot stress how god tier this card and generation was).

The older AMD cards were a lot harder to recommend, GCN4 was just generally slower (but some were dirt cheap so good for budget builds). Vega beats Pascal in performance but only after driver optimization and after it was already released a year after Pascal, it had a lot larger power draw and was priced about the same (very few AIBs existed which sucked as AMD was still doing blower fan cards which were prone to throttling, a liquid cooling version existed but was very pricey). RDNA1 only had midrange cards and while they did offer better performance at times it also lacked any form of RT or AI cores making its future questionable (even though RT and AI acceleration wasn't anywhere near as useful as now, honestly Nvidia was "Skip it" but this kinda felt the same). And then we had RDNA 2 which was absolutely turbulent, at first they looked like decent value cards as their Opencl performance in Blender wasn't that far off from Nvidia with CUDA.
But you might notice a lack of Optix as just like HIP-RT now, optix was crash prone back then. But then we had the bonkers situation with Blender 3.0 and of course the introduction of HIP, no Linux support until Blender 3.2 all the while it loses in about every benchmark in both Cycles and Eevee (+the viewport) even against CUDA. Optix matured by Blender 3.1 widening the gap further and Eevee performance didn't improve until Blender 3.3.

So yeah with Pascal and Ampere I would have recommended Nvidia without a doubt (I owned cards before that like my GTX 770 and Radeon HD 7880 but was less of a number goblin so I really don't know how those generations went) , but the confidence behind that fluctuated as AMD slowly but surely started to improve the situation by software optimization alone. At the release of RDNA2 in 2020 I would have taken the stance "Only if you play games and don't care about ray tracing or anything else" but at the turn of 2023 a lot of "but" would enter my vocabulary around them. Right now AMD would be a clear winner for certain niches, the viewport performance is just mental in terms of price performance, even with last gen cards, and matters when 3D sculpting where materials and lighting take a back seat, for someone who does 3D printing of characters this can be very interesting as you aren't going to optimize or texturing the mesh. The more strokes with dyntopo on, the more FPS will drop and the more you want it to be able to let drop (and not losing to much quality with every remesh). Even Material Preview mode seems to prefer AMD and again in price performance they are just absolutely great value.

I keep forgetting it is a thing and so does the rest of the internet but AMD does have its own Render engine that runs in Blender and even works with Nvidia cards (and has RT-acceleration for both as far as I can tell). It even comes with a free material library. And I saw lightlinking in the feature list which with this patch note it seemed to have even have it before Cycles did. Couple of reviews mentioned that it does seem to interpret settings differently from Cycles, but so does even Eevee and other render engines, I also need to tweak my lighting when I decide I want Cycles over Eevee for a project that I was already actively working on. I might give it a spin to see if it works with grease pencil or not (can't find this documented anywhere) as last time I was hopping through render engines with my GTX 1080 TI, Grease Pencil was a beta feature.

The reddit thread is from 5 months ago, AMD/HIP support in Stablediffusion is again relatively new and ROCM is going through a lot of versions to catch up, the reddit thread mentions ROCM 5.5 but the latest version is 6.0 (mind you it jumped to that on version 5.7 but compatibility and performance has definitely improved since 5.5).
Back then of course this was hard to say if this was going to happen, hell at the release of the 7900XT I would also not recommend it for AI as support was non existent, after ROCM 5.6 and 5.7 it became "Yeah but only if you run Linux, you can't afford top of the line and maybe if you want to train your own LORA" and with ROCM6.0 I can't even tell as I am still waiting on benchmarks but I am excited to finally see competition starting to form.

I just don't want to make the wrong decision when I buy a gpu.

I get that and honestly right now is a pretty turbulent time. A potential for buyer remorse exists as AMD seems to slowly but surely improve their performance with now being a turning point and ZLUDA shows that hardware wise the cards have a lot of fight left in them. For Blender 3.5 the RX 7900XT lost to the RTX 4070 in the whitelands render benchmark, with Blender 3.6 it actually won in that same benchmark. Techgage unfortunately hasn't released a deepdive for Blender 4.0 so I can't say if this remained the case nor can I say what Blender 4.1 might bring on release, Eevee-next might give AMD an edge as it tends to perform better with Vulkan or perhaps the RT-acceleration will aid Nvidia a greater amount as it has better ray tracing performance overall (but Intel not being at the bottom of Eevee benchmarks with all 3 cards is almost a guarantee, they really need Vulkan).

The RTX 3090 is a solid card, bit power hungry. But in my country at least it is well above its launch MSRP and is so close to the RTX 4090 you might as well buy the 4090 (unless you want to barely buy the Ryzen 5 7600 with the 3090).
If you can buy it at the same price as the RX 7900 XTX, it could be very worthwhile since it opens the options to multiple other render engines and has been time tested but if it does cost more keep that in mind when comparing performance and if you are willing to pay that premium, you mentioned Da Vinci Resolve where the RX 7900XTX does have a performance lead over the regular RTX 3090 (with the TI it trades blows), for AI and Blender Cycles the RTX 3090 has better performance as previously shared benchmarks showed, Blender Eevee however is a bit of a depends on the scene (case 1, case 2).

Also, my god man, work on your attitude. If you want information just ask without sounding like a child, don't take a stance you can't defend. And the AMD gpu being recommended for "just gaming" is because a lot of people look at launch reviews, hear that remark (which was true back then) and echo it ever since while never wondering "Is it still true?". Times change, when I first started with Blender when its interface was absolute ass I kept hearing "no studio uses it just learn Maya or 3DS Max (and my god did I hate 3DS Max)" to it being used at Marvel studio in 2018, Ubisoft's animation studio and many more yet you will still hear people echo the age old "Blender isn't used in the industry" (that post is 2 months old) and the further back you set your time range the more you will find more people agreeing with it regardless if it was even still holding true at that point in time.

[Panix]

No need to be insulting tenchrio

I did some research and that's what I base my conclusions/evaluation on - it's not an opinion/conclusion out of thin air. You mention that the reddit post is based on older versions and is 5? months old but then say that optix 'used to crash a lot.'

If you check the Blender opendata website - the Nvidia optix 'tool' improves/increases performance by quite a lot, even lower tier Nvidia cards out perform the top tier of AMD's including the 7900 xt and 7900 xtx. No one has any explanation or rationalization - they can't even say that AMD will eventually 'catch up' in this generation. Q's excuse is that the RDNA 4 will be 'better' (to summarize/simplify). That's good enough - even if true. If AMD is unable to catch up with RDNA 4 - then they're a gaming card series - and the prices should be adjusted downwards, accordingly.

As I said - I received recommendations to go with Nvidia - from ppl in those fields. I even tried to express a potential scenario in which it would make sense for the AMD gpu - Linux use, the potential or possibility that a miracle happens and AMD gets HIP-RT to work - and it's actually somewhat competitive. That the 7900 series is /can be actually good in video editing....that AMD will invest resources and support for other (non-gaming) fields such as AI. It didn't matter - the majority of ppl recommended Nvidia.

The latest techgage site has some benchmarks with Blender and 7900 series of cards vs Nvidia (vs Intel, too). However, Blender 3.6 is tested.

https://techgage.com/article/blender-3-6-performance-deep-dive-gpu-rendering-viewport-performance/

You've probably come across that already?
On the Blender developer website - there is discussion about AMD gpus in Blender - it doesn't sound good to me.

https://devtalk.blender.org/t/cycles-amd-hip-device-feedback/21400/582

There's often talk of AMD gpus not having 'proper tensor cores' or the ray tracing being inferior/lacking - and in the discussion, they talk about 'the extra tools' - HIP-RT - it's there where the major gap happens. It's good that ZLUDA might help close the gap vs CUDA - but, it remains to be seen if there's a similar 'advantage' to be gained vs Optix?
This is recent info (posted):

https://www.reddit.com/r/StableDiffusion/comments/1anov4f/stable_diffusion_for_amd/

https://www.reddit.com/r/StableDiffusion/comments/1agiqg5/what_gpu_to_pick_for_stable_diffusion/

[tenchrio]

No need to be insulting tenchrio

You really do have an attitude problem. And you actually have been insulting from the very get go on topics you seem to have 0 experience with.

I did some research and that's what I base my conclusions/evaluation on - it's not an opinion/conclusion out of thin air. You mention that the reddit post is based on older versions and is 5? months old but then say that optix 'used to crash a lot.'

Yes used to, past tense, before Blender 3.0. You're clearly trolling now. My point was that there was a point in time around the Nvidia's first RTX cards that people still remained skeptical about Optix, things change, Optix is now considered a default settings. Don't drop the "Back then" and pretend as if you are making a point (you are only making a fooI out of yourself).

If you check the Blender opendata website - the Nvidia optix 'tool' improves/increases performance by quite a lot, even lower tier Nvidia cards out perform the top tier of AMD's including the 7900 xt and 7900 xtx. No one has any explanation or rationalization - they can't even say that AMD will eventually 'catch up' in this generation. Q's excuse is that the RDNA 4 will be 'better' (to summarize/simplify). That's good enough - even if true. If AMD is unable to catch up with RDNA 4 - then they're a gaming card series - and the prices should be adjusted downwards, accordingly.

First of all Opendata also has its critique, as Blenchmark used to in the past. You might notice people in that Blenchmark post agree the BMW benchmark is inadequate to measure performance but it is one of the benchmarks still used in the Blender Opendata suite which also used to let you see results per benchmark, now you can't and somehow an RTX 3070 which wouldn't be able to run the Victor Benchmarks with its 9.5VRAM requirement scores better then the RX7900XT (and the RTX 3060 12GB which I linked before beating the RTX 3070 due to the VRAM difference by minutes). There is a reason why tech sites don't use this suite to compare, with Blender 3.6 for example Openbenchmark didn't even give HIP-RT as an option. Not to mention it focuses solely on Cycles, how many times must i repeat this, more Render engines exist, you might value other things outside of just Cycles performance and I can write numerous other reasons why sites user result driven sites like Opendata and Userbenchmarks (the worst benchmark site ever) should not be trusted (a major one being unrealistic/unsustainable overclocking being able to get in between the results).

"No one has any explanation or rationalization" like come on, everyone will say "because of the lack of HIP-RT".
"they're a gaming card series - and the prices should be adjusted downwards, accordingly" This is such a bad take, zluda proves the hardware has a good chunk more potential for other applications, AMD still needs to make a profit and manufacturing costs are a thing, they are already cheaper then Nvidia. The cards do support workloads outside of gaming even if the software isn't always entirely optimized for it and again the extra VRAM wiggleroom is a nice plus as heavy VRAM workloads exist and Nvidia's workstation cards with more vram are more expensive than their gaming cards. And for the millionth time, more Render engines exist with less of a disparity between the two manufacturers and VRAM is important.

As I said - I received recommendations to go with Nvidia - from ppl in those fields. I even tried to express a potential scenario in which it would make sense for the AMD gpu - Linux use, the potential or possibility that a miracle happens and AMD gets HIP-RT to work - and it's actually somewhat competitive. That the 7900 series is /can be actually good in video editing....that AMD will invest resources and support for other (non-gaming) fields such as AI. It didn't matter - the majority of ppl recommended Nvidia.

And I would say wait and am giving my rational and even told you that I was also previously on the same boat as the people you mention and would have recommended Nvidia in the past(especially in 2020 to 2022) but times are clearly changing. If you are so confident in these other people's opinion and don't seem to value mine or the data I present you as clear by the tone of your message, buy Nvidia. Hell I still say it with the RTX 4090, it's pricey but is less likely to disappoint. The rest of the Nvidia 40 series however has a constant problem of price with poor VRAM offerings, AMD's performance is disappointing but sees improvement, even evidence (ZLUDA) that it is still there lurking in the hardware but a lack of VRAM however cannot be fixed, you can't download more RAM no matter what links on the internet you may find. Even in the reddit posts you linked later on people recommend the 3090 over the 4080 and I argued before, it is a good card but it is still pricey, costing more than the 4080 and way more than the 7900XTX, it out performing both would be within the expectation of the price except that scenarios exist were it falls short, the question becomes are those situations applicable or even important to your use case, if you can get it close to the price of a 4080 it would be a great deal but that isn't the case in my country.

The value proposition of extra VRAM is also pretty obvious, considering Nvidia sells their workstation chips at a much higher price despite them sometimes offering 0 improvement in terms of performance even having lower performance. I mentioned it before but look at the RTX 4090 vs the RTX 6000 Ada, the price is almost triple, the performance is less and the VRAM is only double. Either nvidia is out of their mind or this card fits a niche so perfectly that companies are willing to spent 3 times the money just for some extra VRAM. I have mentioned render engines that only work on Nvidia, a company could have bought one of these and are now tied to Nvidia making AMD not an option or they don't use any such render engine but do need over 24GB of VRAM, it is all perspective despite you wishing it wasn't and cherry picking examples while ignoring others I gave you to convince yourself that it isn't.

How many times must I explain that for instance the RX6000 series had a time were it had 0 advantage over Nvidia but now this has changed and it actually has scenarios were it outperforms and provides more value, maybe not Cycles but Blender doesn't stop or start at Cycles. Time isn't static and neither should your opinion be, weigh the current pro's and cons and make your decision from there and unfortunately the scales have been shifting from a clear victor to a case of "but now this aspect is different" and it really depends on how much you value each aspect.

The latest techgage site has some benchmarks with Blender and 7900 series of cards vs Nvidia (vs Intel, too). However, Blender 3.6 is tested.

https://techgage.com/article/blender-3-6-performance-deep-dive-gpu-rendering-viewport-performance/

You've probably come across that already?
On the Blender developer website - there is discussion about AMD gpus in Blender - it doesn't sound good to me.
https://devtalk.blender.org/t/cycles...back/21400/582

Yes, I even linked to some of them previously to point out that RX 7900XT actually managed to win in a cycles benchmark over the RTX 4070 in 3.6(while before in 3.5 the RTX 4070 would be a clear winner every time for Cycles). Are you even paying attention to the links I sent?

And again that is Cycles, even says so in the devtalk you link. Blender is big and complicated and not everything is benchmarked as much as I and other wish it was, 99% of the benchmarks never benchmark fluid simulation which is almost exclusive to CPUs and would make more sense on CPU benchmarks (still I see classroom or god forbid BMW about everywhere) and while GPU accelerated simulators do exist and can be used in Blender, they are still experimental and took a while to develop because crazy thing, fluid simulation is pretty memory hungry and tend to be used to simulate both water and/or smoke in a scene (pretty useful).
To limit Blender to just Cycles is a disservice to how flexible of a tool it is but expected if you never used Blender and don't find out first hand.

There's often talk of AMD gpus not having 'proper tensor cores' or the ray tracing being inferior/lacking - and in the discussion, they talk about 'the extra tools' - HIP-RT - it's there where the major gap happens. It's good that ZLUDA might help close the gap vs CUDA - but, it remains to be seen if there's a similar 'advantage' to be gained vs Optix?
This is recent info (posted):

https://www.reddit.com/r/StableDiffusion/comments/1anov4f/stable_diffusion_for_amd/

https://www.reddit.com/r/StableDiffu...ble_diffusion/

I am not an AI researcher so I can't say if it is true (and neither of the links you provide make that claim, also Tensor is the name of the Nvidia tech, AI cores would be more apt), the AMD RT cores definitely seem inferior but even with Nvidia they don't always aid performance to the same extend in every scenario (I went over this multiple times), it's new tech that only recently became a standard (it definitely wasn't with the RTX 20 Series) so perhaps it might be more consistent in the future.

Also yes that is what I have been saying, the cards themselves have hardware that can offer more performance than they do right now but HIP is still young and it remains to be seen if AMD will fully bring out the potential but regardless there are already scenarios were it is scoring some wins muddying the water compared to 2 years ago. Also both those reddit posts admit the 7900XTX is a good option with Linux (1 again points out that VRAM has importance) and you are posting here on a Linux heavy site, many people sticking to Windows pretend as if Linux is this super hard OS that only someone with a computer PHD could possibly make use of which is one of those things that used to be true perhaps decades ago but now even grandma could easily make use of something like Linux Mint (but they will install Arch and pretend every Linux distro is like it). Perhaps apply that same level of skepticism you are displaying towards me towards them?

You are obviously still fanboying as you refuse to even acknowledge any of the times AMD has better performance, now even linking personal anecdotes of people without any actual numbers to back their claims up (but somehow pretend they are a trustworthy source). Your decision is made, the bias is set, stop pretending as if you are open for alternatives or looking for answers. Go forward with your decision and find out if it is a mistake, all I keep saying is know your use case more and weigh the pros and cons or just go all out on the current king (RTX 4090).

[qarium]

No need to be insulting tenchrio
I did some research and that's what I base my conclusions/evaluation on - it's not an opinion/conclusion out of thin air. You mention that the reddit post is based on older versions and is 5? months old but then say that optix 'used to crash a lot.'
If you check the Blender opendata website - the Nvidia optix 'tool' improves/increases performance by quite a lot, even lower tier Nvidia cards out perform the top tier of AMD's including the 7900 xt and 7900 xtx. No one has any explanation or rationalization - they can't even say that AMD will eventually 'catch up' in this generation. Q's excuse is that the RDNA 4 will be 'better' (to summarize/simplify). That's good enough - even if true. If AMD is unable to catch up with RDNA 4 - then they're a gaming card series - and the prices should be adjusted downwards, accordingly.
As I said - I received recommendations to go with Nvidia - from ppl in those fields. I even tried to express a potential scenario in which it would make sense for the AMD gpu - Linux use, the potential or possibility that a miracle happens and AMD gets HIP-RT to work - and it's actually somewhat competitive. That the 7900 series is /can be actually good in video editing....that AMD will invest resources and support for other (non-gaming) fields such as AI. It didn't matter - the majority of ppl recommended Nvidia.
The latest techgage site has some benchmarks with Blender and 7900 series of cards vs Nvidia (vs Intel, too). However, Blender 3.6 is tested.
You've probably come across that already?
On the Blender developer website - there is discussion about AMD gpus in Blender - it doesn't sound good to me.

https://devtalk.blender.org/t/cycles-amd-hip-device-feedback/21400/582

There's often talk of AMD gpus not having 'proper tensor cores' or the ray tracing being inferior/lacking - and in the discussion, they talk about 'the extra tools' - HIP-RT - it's there where the major gap happens. It's good that ZLUDA might help close the gap vs CUDA - but, it remains to be seen if there's a similar 'advantage' to be gained vs Optix?
This is recent info (posted):

you claim AMD does not have 'proper tensor cores' but see:

They're for machine learning and inference just like Nvidia's Tensor cores. As I mentioned earlier, AI accelerators are parallel matrix math units just like Tensor cores. Just have a different name because branding.
https://community.amd.com/t5/ai/amd-...on/ba-p/637756[/LIST]

and also here:

edit:
OK I stand corrected
RDNA3 indeed seems to have WMMA:

https://gpuopen.com/learn/wmma_on_rdna3/

https://github.com/ROCm/rocWMMA

And I've even found real life application that exceed's AMD's own performance claims:

https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

RDNA3 may not be as bad as it seemed.

"Q's excuse"

I do not have excuses i just say to you what it is and how it will be. for AMD it is simple to put 32GB vram on a 7900GRE
and the RDNA4 8800XT will exactly be such a card.

so in the end your idea that you better go with a 3090 will be wrong thats for sure. more vram is king for the stuff you want to do.

and also Optix is not an argument at all just compare the render results the Optix result is a complete different result than the CPU/CUDA/HIP version.
its not the same result its a complete different picture.

Panix you did not even explain to me if and why the artists would like the complete different result of OptiX...

some maybe like it others would not like it because they want exactly the result of the cpu/CUDA/HIP version

if you don't like the complete different result then why expect people to accept your OptiX argument as a argument at all.-

[WannaBeOCer]

you claim AMD does not have 'proper tensor cores' but see:

To be fair AMD’s RX 7900 XTX performance is equal to the RTX 2080 Ti regarding to machine learning training. They’re not wrong when it comes to Nvidia having better cost to performance when it comes to machine learning.

Nvidia is the easy choice when it comes to a reliable better/cheaper performing platform.

[qarium]

To be fair AMD’s RX 7900 XTX performance is equal to the RTX 2080 Ti regarding to machine learning training. They’re not wrong when it comes to Nvidia having better cost to performance when it comes to machine learning.
Nvidia is the easy choice when it comes to a reliable better/cheaper performing platform.

these benchmarks with the results you claim are not yet useing any WMMA on the AMD side.

because all this software first need to be ported to a version who then accelerate these WMMA units.

[WannaBeOCer]

these benchmarks with the results you claim are not yet useing any WMMA on the AMD side.

because all this software first need to be ported to a version who then accelerate these WMMA units.

They are using WMMA and it’s been implemented in PyTorch.

[qarium]

They are using WMMA and it’s been implemented in PyTorch.

does this mean that you believe that it is fully optimized and can not be optimised more in the future?

[WannaBeOCer]

does this mean that you believe that it is fully optimized and can not be optimised more in the future?

Yes, I don’t believe it will magically overtake the RTX 3080, 3090 let alone touch the RTX 40 series performance. AMD didn’t plan to support ROCM on RDNA3 until they noticed how quickly AI has been taking off. Expect their RDNA4 to have much more than 192 AI accelerators. Their selling point of their AI accelerators on consumer hardware was in regard to future games taking advantage of them.