Mastodawn

@q66 @maya_b ah, big meh! so this is fixed in ampereone?

@mntmn @maya_b i think so; altra max has the same thing (i confirmed it with someone some weeks back) but i believe ampereone is brand new

ofc, it's not available to mortals yet so we won't know for sure for a while

maya_b 🇨🇦Feb 26, 2025

@q66 @mntmn

and at $3k for the least expensive SKU not affordable by most mortals either.

nina

@maya_b @mntmn yeah, though the least expensive sku for that should be beefier than the best one for regular altra, so at least that's somewhat appropriate

but if they messed up the pcie again i won't be amused

elly Feb 26, 2025

@q66 @maya_b @mntmn Ditto on Ampere Altra: Save your time and sanity. It's not worth it.

@elly @q66 @maya_b so disappointing, but thanks for pointing out these issues. boggles the mind how this could be so messed up... i guess we have to wait for an ampereone HPC module then?

@mntmn @elly @maya_b it's a shame because other than this it's an excellent system, with great power efficiency and performance

but if gpus are problematic then it makes a bad workstation (if you do not apply the patches you will get noise in display, if you do you get somewhat slower performance and amdgpu is broken since 6.9 and before then it worked ish but had broken vulkan, funnily what works is nvidia with proprietary drivers with no patches, but you don't want that)

maya_b 🇨🇦Feb 26, 2025

@q66 @mntmn @elly

so okay for headless work but not much else.

@maya_b @mntmn @elly for headless server it's fine yea (unless you need gpu compute, i assume the noise you get in display output unpatched affects even for headless compute in some way)

liffy 💜Feb 26, 2025

@mntmn @elly @maya_b here's what portal2 looked like prior to @q66 's aforementioned patches, by the way! (gdk and qt had to be forced to use software rendering to avoid corruption as well)
as far as vulkan goes, i've had an amount of luck with using AMDVLK when RADV chokes, but yeah...

@lifning @mntmn @elly @maya_b i just get a sigbus with any application using radv on amd

@lifning @elly @maya_b @q66 haha ok that looks interesting

liffy 💜Feb 26, 2025

@mntmn @q66 somewhat interestingly, before the kernel patches, this workaround from something that had previously been encountered on Honeycomb ended up working for this for GL, somehow. (though obviously it wasn't a good long-term solution since the issue turned out to be at the PCI level)

[amdgpu][aarch64] Text and Window Artifacting; Possible System Crash in 3D Applications (#5205) · Issues · Mesa / mesa · GitLab

OS: Linux Ubuntu 21.04 aarch64 System: HoneyComb LX2K ARM64 GPU: AMD Radeon WX 4100 Kernel: 5.11.0.25-generic Mesa Version: 21.3.0-devel (occurs...

GitLab

@lifning @q66 argh, i've encountered that too back when i had a honeycomb...

Val Packett 🧉Feb 26, 2025

@lifning @mntmn @q66 not surprising, honestly i'm surprised no one took these workarounds more seriously

I don't actually remember the context for the RADEON_DOMAIN_GTT and GL_ARB_buffer_storage stuff 0.o but I remember that in FreeBSD we just hardcoded all ioremap to do device memory and added the drm_arch_can_wc_memory() == False patch to Mesa and declared amdgpu to Just Work ™ on the macchiatobin (a8040), honeycomb (lx2160a), etc. (all devices with buggy designware controllers)

@elly @q66 @maya_b ok, one or two more questions: do you have any more detailled info/links about the brokenness level of nvidia/amd/intel gpus on that system? ampere themselves have some guides recommending nvidia rtx cards. are these also broken? because their erratum pcie-65 mentions amd gpus.

@mntmn @elly @maya_b amd: opengl ok before 6.10, vulkan broken; since 6.10, does not work

intel: works ok with the patches since 6.12

nvidia: proprietary driver works fully, no patches (but it's the proprietary driver)

patches reduce performance in edge cases due to alignment fixup

@q66 @elly @maya_b thank you very much for taking the time to list these issues! do you still use the machine, and if so, now with intel GPU? i wonder if sth like blender runs OK on it?

@mntmn @elly @maya_b i use it with an a750, everything more or less works, though it can very occasionally crash the compositor randomly

oh yeah another thing about amd: need to disable amdgpu power management on kernel cmdline or it does not work at all