AMD ROCm 5.5 Released With RDNA3 Improvements, Many Changes

[boxerab]

It's precisely the pace of progress and what outsiders would assume is an underinvestment in resources that frustrates people. ROCm 1.0 was released seven years ago. AMD has a $144B USD market cap. It just feels like this should be going faster.

Exactly. They have the resources to match the nvidia experience for compute on consumer cards, where all cards are supported from day one, on both Linux and Windows. And the public Cuda support forum is fantastic - you will get your question answered by an expert in a matter of hours. I developed an OpenCL application for GCN architecture that ran very fast on Polaris architecture, but they dropped support for this generation in ROCm even though they were selling 580s as late as 2019.
I would encourage anyone interested in compute on consumer cards to port their applications to cuda so they don't have to wait another 7 years for AMD. As the two toolkits are > 90% the same conceptually, it is not hard to create a translation layer that allows the same kernel to run on both.

[WannaBeOCer]

ROCm world is getting more and more interesting ... especially now that Mi50 / Radeon VII cards can be had for peanuts while still offering decent performance and useful amount of memory at respectable memory bandwidth. Perfect entry into the world of ai toys ...

The RTX 20 series can be had for peanuts and runs circles around Vega 20 in regards to AI.

[pegasus]

The RTX 20 series can be had for peanuts and runs circles around Vega 20 in regards to AI.

Maybe ... but ai is not everything.
Which Turing card has 16 or 32gb of memory?
Which Turing card has 1TB/s of memory bandwidth?
Which Turing card has fully open source drivers?

[WannaBeOCer]

Maybe ... but ai is not everything.
Which Turing card has 16 or 32gb of memory?
Which Turing card has 1TB/s of memory bandwidth?
Which Turing card has fully open source drivers?

1. Not required when task are accelerated with fp16, Titan RTX, RTX 5000, 6000 and 8000
2. Again memory bandwidth isn’t needed especially when an architecture has superior compute units.
3. This is just biased

only edge I give Vega 20 is double precision

[Svyatko]

1. Not required when task are accelerated with fp16, Titan RTX, RTX 5000, 6000 and 8000
2. Again memory bandwidth isn’t needed especially when an architecture has superior compute units.
3. This is just biased

only edge I give Vega 20 is double precision

Intel Arc A770 16GB is faster than Radeon VII (including double precision) and costs $350 for new item.
AMD can drop Vega, and GCN5 is not equal to CDNA.
Meanwhile, AMD PAL OpenCL driver still works for GCN2 - GCN5 & RDNA1. IMHO it is possible to use PAL + ROCm libraries (or even Rusticl + ROCm libraries), but ILL no one wants to finance that work.

[WannaBeOCer]

Intel Arc A770 16GB is faster than Radeon VII (including double precision) and costs $350 for new item.
AMD can drop Vega, and GCN5 is not equal to CDNA.
Meanwhile, AMD PAL OpenCL driver still works for GCN2 - GCN5 & RDNA1. IMHO it is possible to use PAL + ROCm libraries (or even Rusticl + ROCm libraries), but ILL no one wants to finance that work.

When it comes to compute the A770 is definitely a better option than Radeon VII for FP16 but Intel's consumer-focused Arc GPUs do not feature hardware-accelerated FP64 cores. For normal rasterization the Radeon VII is faster in regards to games. The A770 is definitely better though since it supports XeSS with dedicated hardware and ray tracing.

Radeon VII(Vega 20) is in a league of its own and probably the last affordable card with double precision.

[dkasak]

I'm waiting on Rusticl. It's embarrassing the BS you have to go through to install ROCm on various Linux distributions. If AMD want sufficient adoption to make their own efforts worthwhile, they really need to focus on this. I've "installed" on Gentoo, Arch, and Fedora, over the past few years. Each time, it's been an unmitigated disaster to my distro's packages database ... and only so I can have the ONLY app which I'm remotely interested in that actually launches ( einsteinathome in the boinc client ) either produce invalid output, or worse - cause a hard lockup. AMD does great hardware. Software ... wow ... not so much.

[Paradigm Shifter]

The discussions regarding pricing of old cards vs. new ones, etc, is amusing to me in a bitter sort of way. A used Radeon VII here will cost >$1000, for a touch more you're in 7900XT territory, and have no warranty or support for the pleasure of that old card which may be dropped from the ROCm support list at a moments notice. And get lucky with a sale and that becomes an XTX. But prices here barely resemble prices in the US/Canada anyway (or even Europe). And some of the shops are well aware of the value of the Radeon VII for anyone wanting "consumer" FP64, and price accordingly.

My attempts to get ROCm to work on not-officially-supported-but-look-people-say-it-works-if-you-do-X-Y-Z cards has been either a) a painful and slightly messy waste of my time or b) an unmitigated disaster so I'm not likely to try again until there is full support on some hardware I don't have to drop $$$$ on just to test. The scenarios where you can get 100%+ performance for ~20% of the price are so rare they just highlight for me how painfully mishandled this has been... for years. I'd settle for (not "never settle", sorry, couldn't resist that dig at the marketing nonsense) for 50% of the performance at 20% of the cost, but right now it's 0% of the performance for any percentage of the cost. And it has been like that for years.

CUDA owns the GPU acceleration segment of my field of research; a system with an nVidia card I can get up and running in an hour or two from not even having an OS (provided speedy internet connection) installed. There are rumours - rumours! - that AMD might be working with a couple of the larger GPU accelerated programs to get ROCm off the ground there, but Intel are already there and the A770 has been available publicly for, what, six months at this point?

I would be more concerned about ROCm, but I've got all excited about positive noises that AMD have made before - multiple times - only to be disappointed, so at this point it just feels like beating a dead horse until long after it is already glue.

And I'm tired of wasting energy on "soon" and "hope to".

And frankly, I think AMD has bigger worries at the moment. While it isn't all their fault (I would argue much of it should be laid squarely on the motherboard manufacturers) the recent stuff about overvolting motherboards and ignoring OCP/OVP protections has actually made me cancel the 7950X3D/AM5 platform purchase I had planned for later this week. It's not urgent, I will watch and wait and observe to see what the resolution is before deciding. That said, I won't be buying Intel instead; I just won't be buying anything.

[NeoMorpheus]

Nobody is ignoring that. You can use CUDA on any green Laptop. Every student can play with their ML models on their private laptop - if they bought the GPU from the competition. If AMD thinks it is not necessary to include those customers, ok. I still can disagree with that decision and bitch about it. Who is gonna start using HIP if you need a dedicated GPU and a specific distro?

In the end, AMD does (somewhat) support gaming GPUs...

Actually, lots of you are ignoring a bunch of facts, when it’s convenient of course.

example, someone complained that ROCm is 7 years old yet ignore that CUDA was launched in 2007. Also complained that AMD market cap is 144B, but ignored that was not the case neither 7 years, less in 2007.

AMD has been fighting against the 2 dirtiest and way more wealthy than them corporations.

But let me be like the hardcore Linux fanatics “you dont like it, then work on the code yourself “ since its clear the world is busting with good developers/programmers that know how to fix this mess in lets say, a month, so all of you would be happy.

I wont say that they havent done mistakes, like depending on OpenCL, which I really don’t understand how it failed, being an open standard with big corporations supporting it.

Anyways, twist all that as you wish, then go buy a nvidia gpu, which is what you really want.😬

[Paradigm Shifter]

AMD has been fighting against the 2 dirtiest and way more wealthy than them corporations.

Yup. And it's for this reason I've been willing to support where possible and I've listened when they've said things were going to happen.

But I'm not made of money, and it's hard to get the bean counters to understand why you want that shiny expensive "compute model" (that's on the officially supported list) GPU for proof-of-concept development when there are already cards available which can do the job, for less money, guaranteed... but with a different colour sticker on the "team" label.

I can, and will, and have in the past, and surely will in the future, spend my own money on (sometimes not so) cheap hardware for proof-of-concept kit. I'd happily buy a NUC-clone APU for ROCm testing (I did, because online people were saying it was possible to get working; I never could, but maybe I bought a wrong model... although I wasn't getting the errors which indicated lack of support for requisites in the UEFI) because I know I can use it for anything else I want if the ROCm adventure proves to be less successful than anticipated (I turned mine into a firewall/proxy). But I'm not dropping ¥270,000 on a W6800 which does not have that flexibility. The APUs are the perfect "consumer" hardware to support to allow low-end (cheap) early ROCm development, even if performance isn't earth shattering, just getting code to run without making my wallet cry is a victory toward leaving Camp CUDA.

But no official support is in sight. Things keep not happening, except at the highest end where I cannot afford entry.

AMD have spent years promising one thing after another. They're obviously capable of making good compute cards, or new supercomputers wouldn't be stuffed full of AMD CPUs and GPUs. But, as others have pointed out, basically the moment CUDA landed, you could run it on any consumer or pro card. Was consumer performance as good? No! But I could buy an 8800GT, and run CUDA code on it on small datasets to prove it would work (although early on CUDA could be difficult to get installed smoothly). I could do the same on a GTX970 - I could actually do real work on 970s (and did!) although the Maxwell Titan gave much more memory breathing room. Or even four 970s, as they were nice and power efficient. And a GTX1080 (in fact, the 1080 was a very, very popular card in my field at the time). I still run a system with a pair of 1080Tis inside for tinkering around with ideas, despite having access to RTX 2080s and 3090's, Quadro RTX 6000s, A4000s, A5000s and A6000s at work.