Radeon ROCm 5.2.3 Released With Ubuntu 20.04.5 Support, Various Library Fixes

[timofonic]

I think they need some heavy shake up in personnel or their hardware support will stay poor forever. Look at this ticket https://github.com/RadeonOpenCompute/ROCm/issues/1400




They are explicitly refusing to provide ROCm / OpenCL support with the same duration of OpenGL / Vulkan to their GPUs. They are only willing to support the current generation HPC supercomputers that they are being paid for. Even Vega is too "old" for them in 2021. Their GPU Compute department don't even recognize that those "embedded" processors are advertised for "5 years" part availability, which means the v1000 mentioned in that ticket is probably still being sold by AMD today. They were removing support recklessly of a product that their company were still selling.

They are making Intel better with that attitude, it's very sad.

I hope they hear us. Where's bridgman from AMD?

[billyswong]

They are making Intel better with that attitude, it's very sad.

I hope they hear us. Where's bridgman from AMD?

And if you don't care "open source" or not, Nvidia is a lot better. From https://developer.nvidia.com/cuda-zone

CUDA for all NVIDIA GPU Families

CUDA serves as a common platform across all NVIDIA GPU families so you can deploy and scale your application across GPU configurations.

Desktop Developer, Data Center Solutions, Embedded Applications, GPU-Accelerate Cloud

Nvidia is listing clearly the hardware support required for a GPU compute ecosystem, in descending order of importance. A Nvidia customer can reliably gain CUDA / OpenCL1.2 support with cards as old as GTX 750 Ti, an 8 years old card, and do GPU compute workload / development on any of them.

[timofonic]

And if you don't care "open source" or not, Nvidia is a lot better. From https://developer.nvidia.com/cuda-zone

Nvidia is listing clearly the hardware support required for a GPU compute ecosystem, in descending order of importance. A Nvidia customer can reliably gain CUDA / OpenCL1.2 support with cards as old as GTX 750 Ti, an 8 years old card, and do GPU compute workload / development on any of them.

That's the sad part. I hope others do the same but Open Source. I dislike AMD attitude a lot.

[qarium]

They are explicitly refusing to provide ROCm / OpenCL support with the same duration of OpenGL / Vulkan to their GPUs. They are only willing to support the current generation HPC supercomputers that they are being paid for. Even Vega is too "old" for them in 2021. Their GPU Compute department don't even recognize that those "embedded" processors are advertised for "5 years" part availability, which means the v1000 mentioned in that ticket is probably still being sold by AMD today. They were removing support recklessly of a product that their company were still selling.

read here: https://www.phoronix.com/forums/foru...alinux-9/page3

vega64 support in Blender 3,3 with ROCm HIP is coming.

OpenCL with image support for vega64 is already done in ROCm

"fedora ~]$ clinfo | grep -i image
Image support Yes"

blender 3.3 on Vega64 with ROCm HIP: https://developer.blender.org/D15242

[qarium]

Vega is too "old" for them in 2021.

AMD is in fact not the relevant driver vendor.

there are already patches for RX480 to RX590 in ROCm and Blender 3.3+ to make it run with HIP...

i say amd is not the relevant driver vendor because amd does not support this in the official driver

but amds driver has very very very low marketshare and if you install ROCm HIP in debian or Arch linux the Polaris/480/580 code is in the driver.

who cares if amd supports it or not? if all the relevant dristro driver vendors support it anyway ?

for example i use this on Fedora:

sudo dnf copr enable mystro256/rocm-opencl
sudo dnf install rocm-opencl
sudo dnf install rocm-*

this means i do not use AMD as a driver vendor at all ...

and all the relevant driver vendors do in fact include the vega and polaris patches.

[Paradigm Shifter]

GARDEN -- General AMD Radeon Development ENvironment

I just had a flashback to Final Fantasy VIII...

[Eirikr1848]

GFX8 also includes GCN 3.0 cards, so not sure why Fiji and Tahiti support is missing. I have a client with a Fiij-based Pro Duo who keeps bugging me to "help him get Blender working right" which is... obviously out of my purview.

That being said there are differences in FP16/FP32 support from GCN 1 and GCN 2 which makes me reluctantly see why those are not included... even though they could be but with reduced performance compared to if they had newer hardware... but even an old data farm is better than no data farm.

As far as money is concerned? Why not be the company that supports older GPUs for longer than 5 years since release? Make it 10 years since the release announcement?

For example someone with 10 cards or 100 cards may want new GPUs... who are they going to choose? The vendor they own now that stops support after a short time, or the vendor that has compute support enabled for a full decade.

10 spins around our star is not long in comparison to the lifetime of... well... anything... but when it comes to a datacenter? Man... I know many a systems engineer or systems admin who would have found a use to cycle the old cards into a less-critical role upon the purchase of new ones... if the older cards had support.

I had to leave actively working thanks to some health issues: but like - other decomm'd server equipment got moved to the "experimental DC" which had older equipment not fit for production: but still had non-prod uses... running side-by-side with ultra-new hardware that did not yet have its place in the main DC.

It seems like someone wants to forget that not all organizations have billion-dollar budgets for a 3-year upgrade cycle. Heck, even enthusiasts and students with a small lab could benefit, and go on to evangelize the brand. "Appeasing the fanboys" so that they promote the product within any organization they work for is like... key Steve Jobs tactic circa 1999/2000 that really worked to get people on-board with Apple products.

TL;DR: If AMD is serious about having a lasting impact in the ML/AI and overall datacenter compute space: it cannot forget about its enthusiasts and students who could go on to be influencers and product purchase decision-makers within larger organizations. Part of that "free" evangelism will come at the cost of ensuring about 10 years of GPU support from launch to software discontinuation. Even then: creating an easy path with documentation for supporting devs to continue product enablement will also be a key to this, so that those with smaller DCs/organizations can easily maintain a mixed-generation environment.

---

Obviously the key here is money: but there are sometimes less... immediately tangible benefits to longer support that shareholders might not immediately see: but if the change was made it would have the chance to improve those investment portfolios well into the future. A teen grows up playing with their parent's rig, inherits an HD 7870. Finds out she can enable Vulkan support unofficially for RADV but then gets confused as to why she can't play around more to set up a working compute dev environment. She then buys an old GTX 750 ti and can install CUDA and play around with its SDK, no issues, and go on to do... remarkable things. She enters a university environment and uses her example to outline why the university should buy nVIDIA products and from there - proceeds to either the research world or corporate world where she climbs the ladder and again evangelizes nvidia products because of multiple decades of great experiences.

Cutting off support because its easy and immediately cost-cutting in some way (frees up a dev, so another does not need to be hired) lacks the foresight of the power of influence and persuasion as it pertains to personal experiences. (Articles and studies written about stuff like this can be googled... but I'm too tired/grumpy from not enough sleep to find them)

AMD is in fact not the relevant driver vendor.

there are already patches for RX480 to RX590 in ROCm and Blender 3.3+ to make it run with HIP...

i say amd is not the relevant driver vendor because amd does not support this in the official driver

but amds driver has very very very low marketshare and if you install ROCm HIP in debian or Arch linux the Polaris/480/580 code is in the driver.

who cares if amd supports it or not? if all the relevant dristro driver vendors support it anyway ?

for example i use this on Fedora:

sudo dnf copr enable mystro256/rocm-opencl
sudo dnf install rocm-opencl
sudo dnf install rocm-*

this means i do not use AMD as a driver vendor at all ...

and all the relevant driver vendors do in fact include the vega and polaris patches.