AMD ROCm 5.5 Released With RDNA3 Improvements, Many Changes

[pegasus]

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,

And most powerful supercomputers (AMD based) run on rumors? Think again.
IMHO those big machines will help AMD get their sw stack into shape and that will benefit all the rest of us as well. Some of us just want to get the excitement of being early adopters, some don't care and just use what works for them. Both approaches are perfectly valid.

[Grinness]

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...

While AMD support of laptop GPU is not good, dGPU hardware requirements is pretty much aligned across the 2 vendors
For example, to use Pytorch on a 'green card' you need Cuda 6 (GTX 10 series +):

https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel_19-02.html

Same era of GPU as the 'red cards' (Polaris)
I have been running ROCm + pytorch since Polaris (rx 480 8GB)

With regards to distro, you can get ROCM on (essentially) all main distros and derivatives using pre-compiled packages:
* rpm (Fedora, Redhat, ...)
* Debian (Ubuntu, ...)
* Arch (Manjaro, ...) -- ROCM packages are in Arch official repos since 5.3 or 5.4

https://github.com/RadeonOpenCompute/ROCm_Documentation/blob/master/Installation_Guide/List-of-ROCm-Packages-for-Ubuntu-Fedora.rst

https://archlinux.org/packages/?q=rocm

[CochainComplex]

Even if the thread is not representative for the whole market it gives a glance at the current situation.

People are willing to use ROCm dispite the overwhelming domination of CUDA. But the reason why people are not trying it or getting frustrated is because there is lack of support of "try me" opertunities.
If you want to convince people to switch, you need to make it as easy and as convenient as possible. Especially if there is already an intrinsic motivation. Those are the easiest people to sell the product to. But if you fail them, something is wrong with the selling arguments and your approach.

Giving people the advise: yeah if you want try it, go to eBay and find a decent old pro card for a "cheap" price and install it in a dedicated PC with a clean distro setup such that you have no other packages interfering with ROCm...can't be the selling argument.

Why not supporting APUs with "abysmal" AI or Computing performance if people are tinkering and publishing their project on git where maybe thousands of people with equally weak hardware are going to use it? This is how you break dominant market leaders. By makeing people aware of well working and convenient alternatives. (Real and not hypothetical) Maybe one of the thousand users needs exactly this project and wants to scale it bigger for professional purposes? Is this so unlikely?

If I want to try a new brand of cars the brand does not recommend me to go and find a second hand car for a "good price" to check the brands quality and capabilities. You go to the brand distributor and ask for a testride. Easy and convenient.

(Recommending second hand products for the first hands experience can also fire back since potential issues can have their cause by the previous owner and not by the brand it self)

How did AMD convince Linux Gamers? Exactly - with a driver architecture which does not break by switching kernel or packages and runs on any potato and is installed wihtout any notice. Convenience.

[Paradigm Shifter]

And most powerful supercomputers (AMD based) run on rumors? Think again.

I meant rumours explicitly within my field of research. Which was clear in the original text before you removed that sentence from the paragraph which gave it context. Read again, then think again.

IMHO those big machines will help AMD get their sw stack into shape and that will benefit all the rest of us as well. Some of us just want to get the excitement of being early adopters, some don't care and just use what works for them. Both approaches are perfectly valid.

On this we agree. But getting their software stack in order on one or two supercomputers might not be the best strategy to follow for wider adoption at a reasonable pace.

[CochainComplex]

pegasus It is nice that the most powerfull super computers run on AMD. But how many real users are using it? At my work we have a huge HPC it was once top1 on the green IT Supercomputer list. But do you know how many frequent users are using it? Max 100 people (maybe 200 on-off base). Considering the size and the cost of this computer this is nothing.
A lecture hall filled with students and their laptops have more CUDA users.

I have a Nextcloud instance running on an AMD APU based micro server. Recently I have figured out that Nextcloud has a built-in function * which sorts and adds lables to pictures with support of AI. it makes use of pytorch....unfortunately I can only use it with the CPU.
After 2,5 weeks of constant 100% load on the CPU I have stopped the software. Only half of my pictures got labled.
For a glimps of second I was considering to buy a potato Nvidia Card with CUDA support.
Missed opportunity of AMD to show me that ROCm is a serious alternative. Even if I hate Nvidia for its greedy buisness tactics I was considering for a second to get one of their products.

* Photos - https://github.com/matiasdelellis/fa...nd-Limitations ..1-2 sec GPU ...40-120 sec just CPU

[CochainComplex]

maybe this thread might be interessting to bridgman

[CochainComplex]

btw you might find APU support here / but might be a hard ride:

https://bruhnspace.com/en/bruhnspace-rocm-for-amd-apus/

[Svyatko]

btw you might find APU support here / but might be a hard ride:

https://bruhnspace.com/en/bruhnspace-rocm-for-amd-apus/

They'd stopped at ROCm 3.10 https://www.phoronix.com/news/Radeon-ROCm-3.10-Released
Possibly no OpenCL image support. Troubles with modern releases. Possibly no RDNA/RDNA2 suport.

AMD could finance Bruhnspace, but decided not doing that
In 2020 and later AMD had enough money, but preferred to cancel investments in GPGPU future.
Possibly this is because of lack of strategic planning.

[Svyatko]

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.

https://www.reddit.com/r/learnmachinelearning/comments/xxozea/tensorflow_on_intel_arc_gpu/

https://www.reddit.com/r/learnmachinelearning/comments/xxozea/comment/iw2g42k/?utm_source=share&utm_medium=web3x&utm_name=web3xc ss&utm_term=1&utm_content=share_button

I gave Intel Extension for Pytorch (a.k.a. IPEX) a shot using my i5 11400H's integrated graphics (yes IPEX can run on basically any Intel GPU that oneAPI supports which goes as far back as Skylake iGPUs as listed in Intel's documentation here), and I would highly NOT recommend using IPEX NOT because of performance issue reasons (I didn't expect my integrated graphics to be fast anyway for obvious reasons), but because I feel that Intel basically shot themselves in the foot with the lack of FP64 instructions hardware support in Gen11 and onwards graphics chips (e.g. Tiger Lake, Alder Lake, Arc Alchemist, etc.), as it is impossible to not only run any kind of workload involving FP64 instructions using IPEX given how FP64 emulation is an utterly broken mess as of the time of me writing this, but there is absolutely zero documentation right now on how to write python scripts that one can run using IPEX without needing to use FP64 emulation. So yea until Intel either puts out concrete documentation on how to write scripts that doesn't require use of FP64 emulation or just straight up fix their broken FP64 emulation, I might end up having to buy an Nvidia GPU for my machine learning needs.

https://www.reddit.com/r/learnmachinelearning/comments/xxozea/comment/j9nvq77/?utm_name=web3xcss

Funny you should ask - I spoke with an Intel engineer on GitHub about how I considered getting an Arc for both pytorch (i.e. use with Intel extension for pytorch) and tensorflow (i.e. use with Intel extension for tensorflow), and I was basically told to not waste my hard earned money on something that was only validated for Intel data center cards and not Intel consumer cards lol, so I just ended up getting an Ampere Quadro instead 🤣🤣🤣

Sorry mate to opt out of the Intel beta testing crew; when I get told by an Intel engineer to basically not buy an Intel product I'd rather follow the engineer's advice rather than risk making a decision that I might regret later. 👀

So sad...

[Mahboi]

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.

You just described HIP.

Also this is the fundamental problem with AMD, great hardware, grossly insufficient support of it compared to the competition. The excuse of "being smaller" runs thin as it's been years since AMD has been making great bucks. Their 2020 revenue was $9B, their 2022 revenue was over $20B. The whole "we are too small to really compete, plz accept 20% cheaper for 50% of the support" is becoming more and more of a stretch every day. In CPU it's a minor/non issue, in GPUs, it's really losing any kind of patience.