Open-Source AMD Radeon Linux Graphics In Great Shape For Workstations, Handily Beating Proprietary Driver

[lyamc]

Blackmagic needs to get their head out of their butts with Davinci Resolve

[david-nk]

Mesa performance for Radeon cards seems to have evolved more in recent years than I expected.
Are there are any benchmarks for Mesa 20/21/22 on Phoronix? There probably are, but I couldn't find anything using the search function.

[Linuxxx]

Michael

Really interesting benchmark & results, thanks for that!

Just one minor nitpick:
Maybe using the performance governor instead of ondemand would have pushed the open-source difference even further ahead of the proprietary competition because of the lower CPU overhead.
And since we are talking about workstation loads here, I assume a properly configured one should make use of the performance governor all the time.

(BTW, would be funny to know how many companies are spending more than 10k $ per workstation, only then to install an enterprise Linux distro with default settings & thus leaving alot of CPU power on the table...)

[oiaohm]

There are two separate teams. We still need to support other OSes so not having a closed source driver on Linux does not suddenly free up a lot of resources to work on Linux. Windows still needs a driver.

But there still should be a question since mesa3d is under MIT class license you can use mesa3d in the closed source drivers. Would it be worth the cost dropping the Linux closed source driver and putting resources in to have mesa3d be functional on windows by like implementing GL_EXT_memory_object_win32 and GL_EXT_semaphore_win32.

Yes Windows still needs a driver but that does not say that driver has to be 100 percent closed source. Yes this could possible reduce the number of teams. Or at least reduce duplication in effort.

Yes serous question currently not answers is the open source implementation only better under Linux or is it better everywhere. Of course to find out this mesa3d and other parts need to be taken to other platforms with drivers to access the gpu fully.

Yes it would be as interesting world if Intel and AMD was in fact sharing gpu driver development work on all platforms not just open source ones.

[zexelon]

Neat. I wonder if AMD could eventually use some of the RadeonSI code in Windows?


At the risk of going off topic, I'd like to see some ROCM benchmarks from Phoronix if possible, ideally against an Nvidia RTX card running the proprietary drivers. PyTorch/TF performance will be a huge factor in my own buying decisions. The last time I had a AMD dGPU, I ragequit just trying to set rocm up, but it seems that the packaging has gotten *much* better now.

I second this, I want to see rocm vs CUDA... its about the only performance metric that matters to me these days. Would love to give AMD another shot at GPU but cant without a real CUDA replacement.

[Mystro256]

2 teams, gotta ve honest, that's the dumbest fucking thing I've read in a while.

Just go to only open source and internally build only with the open source driver.

You've obviously never seen the mess that is Windows driver development. Sure Mesa would work on Windows, but there's piles of things that MS needs *cough*DRM*cough* that makes mesa pretty challenging to use. Although, Mesa is MIT I think, so they might be able to fork it and try to carry some non-free patches on a private tree, but it definitely doesn't come for free.

[onlyLinuxLuvUBack]

Could you do a part2 of this article where you run (clear linux and then distrobox and then this same ubuntu under distrobox) to see if perf loss or gain ?
and also compare in graph: (fedora silverblue + distrobox + this ubuntu) and (newest alma + distrobox + this ubuntu) ?

[brucethemoose]

You've obviously never seen the mess that is Windows driver development. Sure Mesa would work on Windows, but there's piles of things that MS needs *cough*DRM*cough* that makes mesa pretty challenging to use. Although, Mesa is MIT I think, so they might be able to fork it and try to carry some non-free patches on a private tree, but it definitely doesn't come for free.

Yeah, a private fork is exactly what I was thinking. They can still upstream fixes and improvements, if only so their own fork doesn't get too out of sync.

Imagine if even a quarter of the work put into the closed driver was instead put into Mesa.

[obri]

In fairness, Marek put a lot of work into improving workstation app performance in the open source driver over the last couple of years. Without that I suspect AMDGPU-PRO might have been faster. Either way it's good to see.

Do I understand you right, that Marek is an AMD dev working on the proprietary driver who helps also in developing the opensource driver?

[obri]

There are two separate teams. We still need to support other OSes so not having a closed source driver on Linux does not suddenly free up a lot of resources to work on Linux. Windows still needs a driver.

Is it possible to do something like opensourcing the Windows driver and sharing the oss codebase between Win and Linux? AFAI understand, actually the closed source Linux driver shares the codebase with the Windows driver.
Perhaps it is completely naive, but wouldn't it save ressources and bring more speed to the driver development, when you would do it with one driver oss code base, and work as a team on one driver then?

From the outside it may seem as a good idea, but of course I do not know your details and dependencies and management requierements and so on.

As a 6800 XT owner and Linux user, I would naturally be happy with as much quality and functionality in the oss driver such as good raytraycing for example.