AMD Releases ROCm 6.0.2 With Improved Stability For Instinct MI300 Series

[phoronix]

Phoronix: AMD Releases ROCm 6.0.2 With Improved Stability For Instinct MI300 Series

AMD on Wednesday evening released ROCm 6.0.2 as the newest point release to their open-source compute stack...

https://www.phoronix.com/news/AMD-ROCm-6.0.2-MI300

[Danny3]

Is it really open source?
If it is, why aren't Linux distros come with it installed by default, like Mesa?
Why do we still have to jump through hoops to have compute working?

[s_j_newbury]

Is it really open source?
If it is, why aren't Linux distros come with it installed by default, like Mesa?
Why do we still have to jump through hoops to have compute working?

That's a good question. Since I'm trying to maintain a Gentoo overlay for ROCm, I may have some insights:

First of all, ROCm may or may not work with upstream LLVM in any given release, or break with LLVM updates. There is a separate AMD LLVM fork which should work with a particular release of ROCm, but distros won't want to have to package duplicate LLVM toolchains.

Coverage of GPU support is an issue. Despite AMD stating here recently that "unsupported" GPUs (which did work in many cases, or were officially supported) should still be able to be used, such as Vega56/64 (gfx900), this doesn't seem to be communicated to all sub-projects, which results in the code being removed, or at least deliberately prevented from being built. This makes it hard for distributions to communicate to users whether their hardware should work, of whether it will break with an update. This puts all the burden on any potential distributions to keep support for GPUs working or be flooded with bug reports.

While the OpenCL support from ROCm has excellent OpenCL 2 support, for me it's currently broken in my latest builds (for gfx900, see above) and I'm trying to figure out why, maybe 6.0.2 is better... Hopefully Rusticle will flesh out decent OpenCL 2 support in the near future, which won't help incentivise anybody to package ROCm.

AMD HIP just isn't that well supported by projects, in comparison to CUDA. Many of them could support ROCm by using hipify to convert their CUDA sources, but there isn't the demand due to how few people have working installs of ROCm. It's a "Chicken and Egg" situation. Distributions aren't going to go to all the effort if there isn't anything which actually uses ROCm. This also relates to the perception and reality of what GPUs actually work with ROCm as far as whether any particular project will even consider the extra work.

[darkbasic]

I'm trying to maintain a Gentoo overlay for ROCm

Why don't you contribute to the official rocm ebuilds instead? I've done so myself to try getting rocm working on ppc64le (unsuccessfully) and they merge PRs fairly quickly.

[s_j_newbury]

Why don't you contribute to the official rocm ebuilds instead? I've done so myself to try getting rocm working on ppc64le (unsuccessfully) and they merge PRs fairly quickly.

The real reason, or my excuse? My excuse is that it allows me more flexibility to try things out such as using flang(-new), or AMDs LLVM fork. The real reason is because I got tired of dealing with other developers, I spend most of my time fixing up other peoples breakage or removal of things I depended upon, and sometimes my head gets sore from wall banging. I used to be very active on the Gentoo Bugzilla. If you say the Gentoo ROCm maintainers are okay to deal with then maybe I'll give it a go, it does make more sense to collaborate.

[Panix]

Is it really open source?
If it is, why aren't Linux distros come with it installed by default, like Mesa?
Why do we still have to jump through hoops to have compute working?

AMD doesn't care about Linux or GPU Compute. They care about AI - but as it pertains to Windows.

[Panix]

Why don't you contribute to the official rocm ebuilds instead? I've done so myself to try getting rocm working on ppc64le (unsuccessfully) and they merge PRs fairly quickly.

Why doesn't AMD support their products? Their shoddy Linux support combined with little support in GPU Compute should be obvious even to AMD fanboys.

[Panix]

Here's something for the brainwashed, idiotic AMD fanboys on here:

https://community.amd.com/t5/rocm/ct-p/amd-rocm

AMD users crying that ROCm doesn't work.

https://www.reddit.com/r/ROCm/comments/18jrzzc/rocm_600_release/

It sounds like only two cards are supported (officially). LOL! AMD DOESN'T SUPPORT THEIR PRODUCTS!

I'm interested in this area myself - somewhat - I am more interested in Blender and video editing but I keep track of these developments. In EVERY area, AMD is not only behind, but their support is shoddy. Their MAIN AREAS OF INTEREST/FOCUS is in gaming and AI. I guess they need ROCm for the AI part so that will be another WIP and maybe they won't take a million years to get it implemented/working this time.

Other than their processors and strictly gaming, they're really incompetent and they're just as greedy as Ngreedia - they are both crappy companies and neither should have 'fanboys.'

[s_j_newbury]

Here's something for the brainwashed, idiotic AMD fanboys on here:

https://community.amd.com/t5/rocm/ct-p/amd-rocm

AMD users crying that ROCm doesn't work.

https://www.reddit.com/r/ROCm/comments/18jrzzc/rocm_600_release/

It sounds like only two cards are supported (officially). LOL! AMD DOESN'T SUPPORT THEIR PRODUCTS!

I'm interested in this area myself - somewhat - I am more interested in Blender and video editing but I keep track of these developments. In EVERY area, AMD is not only behind, but their support is shoddy. Their MAIN AREAS OF INTEREST/FOCUS is in gaming and AI. I guess they need ROCm for the AI part so that will be another WIP and maybe they won't take a million years to get it implemented/working this time.

Other than their processors and strictly gaming, they're really incompetent and they're just as greedy as Ngreedia - they are both crappy companies and neither should have 'fanboys.'

I don't think they really are all that far behind with respect to "Compute", there is though a community engagement problem.

Gaming only works well because the community (Mesa devs) did the heavy lifting, AMD really got on board and made it work well after it was clearly going to leave their proprietary FireGL driver in the dust. Yet they still didn't get on board with RADV initially, pushing their own in-house Vulkan driver.

ROCm, whilst "Open Source" isn't a community project, this is clear by the way they actively deprecate and remove working GPU support, yes this happens in the Linux Kernel, and Mesa too, after several decades when nobody steps up to maintain the drivers! [Let's not get into the recent removal of IA64 here!] Removing support for GPUs like the RX Vega cards isn't something a community project would do.

If AMD wants to develop a community around ROCm they need to consider community needs and reach out to people like myself who are prepared to do the work of maintaining support for hardware they no longer prioritize. They then can promote their compute stack of various distributions as "supporting" various models beyond what they themselves focus on. They don't need to "support" a binary release for every distribution out there, or any at all if they can get distributions to do the integration and validation. It just means they need to stop putting up roadblocks and discouraging contributors.

[Panix]

I don't think they really are all that far behind with respect to "Compute", there is though a community engagement problem.

Gaming only works well because the community (Mesa devs) did the heavy lifting, AMD really got on board and made it work well after it was clearly going to leave their proprietary FireGL driver in the dust. Yet they still didn't get on board with RADV initially, pushing their own in-house Vulkan driver.

ROCm, whilst "Open Source" isn't a community project, this is clear by the way they actively deprecate and remove working GPU support, yes this happens in the Linux Kernel, and Mesa too, after several decades when nobody steps up to maintain the drivers! [Let's not get into the recent removal of IA64 here!] Removing support for GPUs like the RX Vega cards isn't something a community project would do.

If AMD wants to develop a community around ROCm they need to consider community needs and reach out to people like myself who are prepared to do the work of maintaining support for hardware they no longer prioritize. They then can promote their compute stack of various distributions as "supporting" various models beyond what they themselves focus on. They don't need to "support" a binary release for every distribution out there, or any at all if they can get distributions to do the integration and validation. It just means they need to stop putting up roadblocks and discouraging contributors.

AMD is really behind (Nvidia).

I was interested in AMD gpus - I am annoyed because they're so far behind - is there anyone here who isn't an AMD fanboy who would recommend an AMD gpu like a 7900 XTX or XT in Linux if using Compute? I don't think so/I doubt it. Ppl buy Nvidia gpus for that - even in Linux.