Radeon ROCm 5.4.1 Released

[NeoMorpheus]

Not that i agree with AMD actions entirely here, but I remember that currently, they have 2 distinct architectures:

1- Compute-heavy units (CDNA) based gpus for professional use.

2- Rasterization-heavy units (RDNA) based gpus for gaming.

Hence why the clear divide. That said, i dont know if that its smart on their part, but its clear that people dont agree with it.

Lastly, seems like they are increasing the support for rdna devices, but i wonder if the necessary hardware for this task is there at all.

[Alpha_Lyrae]

Not that i agree with AMD actions entirely here, but I remember that currently, they have 2 distinct architectures:

1- Compute-heavy units (CDNA) based gpus for professional use.

2- Rasterization-heavy units (RDNA) based gpus for gaming.

Hence why the clear divide. That said, i dont know if that its smart on their part, but its clear that people dont agree with it.

Lastly, seems like they are increasing the support for rdna devices, but i wonder if the necessary hardware for this task is there at all.

RDNA is a compute architecture with graphics and RDNA3 expands that compute with selective 2xFP32 capability. IMO, it’s actually a better compute platform for branchy code than CDNA due to wider SIMD32 and 1-cycle wave32 capability (curious that an Nvidia warp is also 32-thread) or in RDNA3, 1-cycle wave64 capability. CDNA is pure compute with media codecs and no graphics hardware at all. Much more datacenter accelerator, as it also has no display engines. So, RDNA has to fill the in-between spaces for professionals that need graphics and display outputs.

RDNA, for all intents and purposes, is rebranded GCN 2.0, where Tahiti to Vega were 1.0-1.5. CDNA is still carrying GCN from where Vega left off and is still 4-cycle SIMD16 and only wave64.

[finalzone]

Installing ROCm OpenCL and HIP on this release from AMD website for say Fedora is much easier than before. Good to see both RX 6950 XT and the the gpu part of Cezanne APU is supported. It will be a matter of time dependencies get ironed out.

[Spacefish]

RDNA is a compute architecture with graphics and RDNA3 expands that compute with selective 2xFP32 capability. IMO, it’s actually a better compute platform for branchy code than CDNA due to wider SIMD32 and 1-cycle wave32 capability

RDNA3 might be more suitable for machine learning tasks (RocmMI) than RDNA2/1, as it adds Matrix cores which support bf16 and WMMA.. This was a feature exclusive to CDNA before.. AMD didn´t really bother so support / optimize most machine learning kernels on RNDA1/2 as far as i understand it, as the performance would be pretty lacking compared to CDNAs Matrix cores.
Still recent CDNA cards (Mi200) have way more matrix cores + way faster memory subsystem compared to the RDNAx counterparts.

Especially with more AI / machine learning and raytracing now in consumer applications the architectual gap between CDNA and RDNA closes a littlebit. Maybe they will in the future with a new generation switch back to a single core design / modual core design, where more of the logic is shared? No one knows..

[marlock]

IMHO if AMD wants people to grow into data science roles and other compute-heavy jobs using their hardware and an open stack instead of just maintaining the Nvidia+CUDA status-quo, they need to support compute flawlessly in non-PRO GPUs out-of-the-box ASAP...

My reasons to say this:
- Nearly nobody buys professional-grade GPUs for their home computers, even though they might buy 64GB+ of RAM and a Ryzen 7 5xxx or even a Threadripper CPU. It's not that big of a price jump to invest in, and there is no downside to a better RAM or better CPU for other uses like there is for gaming on a PRO GPU.
- Not all such jobs require such a large-scale processing that justifies a special workstation-grade machine... We're undergoing a full-on digital transformation process in many companies, so people are dipping their toes in data science, machine learning, compute and etc en-masse... and they'll do it on their current plain machines BEFORE deciding if it's worth an upgrade.
- Laptops are much more frequently the only machine promptly available in some companies and jobs, so iGPUs should be usable for such things... and what-do-ya-know, Ryzen iGPUs are the best thing available on laptops right now save insanely-priced dedicated GPU hardware or workstation-grade laptops weighting more than a brick wall.
- Some tasks might need development / prototyping before deployment, and plain, portable and/or home machines are frequently enough for that (and might even offer a more comfortable environment to work with before deployment... and now we have a much higher ratio of home-office / hybrid workers so this is not going away anytime soon.
- New GPUs are forbidingly expensive even without the price premium of being a professional-grade model, and not just people but even serious companies, universities and etc are short on cash... if entry-point AMD hardware could be sold now then upgraded later, wouldn't that be better than keeping all that marketshare on the green side as it's been for a while?
- OpenCL acceleration can be used for daily-use plain software too! LibreOffice has an optional method for this that could be enabled by default if only OpenCL would run properly everywhere...
- Even gaming is now turning thirsty for compute/ML, so the divide is melting away

I'm sure part of the reason AMD is lagging behind is historical, with them almost going bankrupt at some point it's natural that some things got pushed aside for a while... but AFAIK they're safely out of the red zone now and I hope they can pick up the pace on this front as they've been doing on others.

[LtdJorge]

IMHO if AMD wants people to grow into data science roles and other compute-heavy jobs using their hardware and an open stack instead of just maintaining the Nvidia+CUDA status-quo, they need to support compute flawlessly in non-PRO GPUs out-of-the-box ASAP...

My reasons to say this:
- Nearly nobody buys professional-grade GPUs for their home computers, even though they might buy 64GB+ of RAM and a Ryzen 7 5xxx or even a Threadripper CPU. It's not that big of a price jump to invest in, and there is no downside to a better RAM or better CPU for other uses like there is for gaming on a PRO GPU.
- Not all such jobs require such a large-scale processing that justifies a special workstation-grade machine... We're undergoing a full-on digital transformation process in many companies, so people are dipping their toes in data science, machine learning, compute and etc en-masse... and they'll do it on their current plain machines BEFORE deciding if it's worth an upgrade.
- Laptops are much more frequently the only machine promptly available in some companies and jobs, so iGPUs should be usable for such things... and what-do-ya-know, Ryzen iGPUs are the best thing available on laptops right now save insanely-priced dedicated GPU hardware or workstation-grade laptops weighting more than a brick wall.
- Some tasks might need development / prototyping before deployment, and plain, portable and/or home machines are frequently enough for that (and might even offer a more comfortable environment to work with before deployment... and now we have a much higher ratio of home-office / hybrid workers so this is not going away anytime soon.
- New GPUs are forbidingly expensive even without the price premium of being a professional-grade model, and not just people but even serious companies, universities and etc are short on cash... if entry-point AMD hardware could be sold now then upgraded later, wouldn't that be better than keeping all that marketshare on the green side as it's been for a while?
- OpenCL acceleration can be used for daily-use plain software too! LibreOffice has an optional method for this that could be enabled by default if only OpenCL would run properly everywhere...
- Even gaming is now turning thirsty for compute/ML, so the divide is melting away

I'm sure part of the reason AMD is lagging behind is historical, with them almost going bankrupt at some point it's natural that some things got pushed aside for a while... but AFAIK they're safely out of the red zone now and I hope they can pick up the pace on this front as they've been doing on others.

They of course know all this, but you can't build CUDA 2 in a year. They first focused on HPC, which brings in the big money in comparison.

[marlock]

You're right, it's obviously expected that they focus on HPC first, except I was talking about OpenCL, not some new derivate HIP solution... it's been here for ages, even ROCm exists since ages... why don't they fix the spotty hardware support (to say the least) before piling up even more features only a few people will risk using? It all sounds great, but how can anyone trust it's gonna work on their machine if there is such a sea of bad feedback? Where is the backup army to speed up their development and get to a more profitable scenario when they have a chance?

PS: I'm not mad at them, just puzzled...

PPS: they'll now have a second window of oportunity in Intel's OneAPI, which was recently released with AMD support on top on ROCm and Nvidia support on top of CUDA, but honestly I'm not holding my breath that AMD themselves are going to work hard to get it in good shape for their hardware... it's much more likely to be done by RedHat, Collabora or some other FOSS player in the Linux space

[LtdJorge]

Sorry, I meant they went for HPC hardware when building compute. The focus on compute for consumer hardware comes later.

Re: OpenCL, at least they have a decent implementation, unlike Nvidia which neglects it so no one bothers to use it. Sigh...

[marlock]

Compatibility layers are the new black

It's the only way to do other people's work for them in a way they can't stop you from doing it.

Each busted walled garden makes me happier and happier

eg: using CUDA to run OneAPI on Nvidia GPUs from day one was a genious move from Intel in that sense.