A Look Back: When Everyone Had Problems With ATI/AMD On Linux

[AnonymousCoward]

opencl rendering

Interesting, never heard about that. What's the reasoning behind it? or why doesn't it work on AMD cards?

[curaga]

Oh those days. My first and for a long time last AMD blob experience was trying the last supported blob for the Radeon 9250. The screen was way overbright, couldn't see shit, and it crashed X after logging in. I also had some i810 hw, which was absolutely horrible in every OS, random bugs in Windows too.

So, I used Matrox with excellent open drivers for a few years, then moved to stabilized FOSS radeon, still there. AMD's blob hasn't improved much from that state.

[Nille]

Interesting, never heard about that. What's the reasoning behind it? or why doesn't it work on AMD cards?

Then you Never looked at this. The Kernels are Valiv but amds implementation dont like the size and cant compile it. If you want to use opencl in blender you has to turn off most of the features.

[YoungManKlaus]

7 years ago

my brain: "jeah, that must have been somewhen around 2000, right?"

[Sdar]

Interesting, never heard about that. What's the reasoning behind it? or why doesn't it work on AMD cards?

Just what Nille said... but i want to add some information.

Blender Cycles render using Opencl works with intel and Nvidia opencl compiler but not with AMD compiler cause some bugs, this was reported like ~3 years ago and AMD stated that it's a bug and they're working on it, they even said that it was a high priority a couple of times... then they said that it'll be fixed soon but only for GCN cards ,like a year ago, and suggest everybody to renew their cards to the new GCN architecture promising news soon for this cards, then they said that they're going to fix it only for RadeonPro GCN cards... then they said that they already have a fix but it didn't pass the QA tests...

So we've been hearing that it'll be fixed soon for the past ~3-4 years... while it works with intel and nvidia implementation, it has problem even with luxrender where you can't render if the scene uses too many features at the same time, also when a kernel is large but it still works the performance drop is huge causing low end Nvidia cards to surpass high end amd cards performance in some cases.

Even the CPU opencl compiler is broken: http://www.luxrender.net/forum/viewt...p?f=34&t=11009

So to get a reliable Opencl compiler you need to use Nvidia even if it's just Opencl 1.1 compatible is still better than AMD Opencl 2.0 compiler.

[droidhacker]

Ehrm... Chris DiBona's prediction cannot be shown to be incorrect, just because the process takes longer than you anticipated.
As it stands, a much greater proportion of graphics hardware TODAY is supported by open source drivers (to varying degrees) than was the case 7 years ago.

And for that matter, even Satan/Nvidia have shifted somewhat. They were *incredibly* hostile to open source back then, but more recently, I hear that they have actually made some positive contributions to nouveau, and been looking into open source for their *mobile* hardware. Granted, it is not nearly enough for me to give them a shot, but it sure is different than it was several years ago....

[opensource]

What doesn't work? I used it in the past and present on two different AMD cards with both open and closed drivers and never had any issues. We are talking about the same Blender, right?

Did you use the cards for rendering? Because Blender Cycles doesn't work (it may be able to render but will render incorrectly and even if it will render (incorrectly), it will render at only ~1-2% speed, so very very slow) -- a three years old bug. When I bought my AMD card, I didn't know that, but now I'll sell it and buy a NVIDIA card, even if it's OpenCL 1.1, not 1.2 or even 2.0. Since Maxwell 2.0 (GTX 970,980) watt TDP is very good, Hashcat (integer speed) is very good and LuxMark 2.0 benchmarks are also at the top.

[blackiwid]

- Chris DiBona thought closed-source Linux GPUs would collapse if one of the major GPU vendors open-sourced their driver and specs. Well, AMD did it, but his prediction was wrong

The question is what u understand with taht part, make a new one or opensource their blob. Catalysts design maybe is not best, but people would have fixed the bugs and even when people would have still made a mesa driver they could always look in the code and see exactly how it would have worked, not waiting 5 years till uvd doku/driver gets released.

The last doku drops are not that old maybe maximum 2 years old. So if they really would have opensourced their driver, not made a new one r600 and now radeonsi his prediction would be more likely become true. So I dont think he predicted wrong, just it did not happen, there is a AND between driver and specs no OR.

[Calinou]

Like 7 years ago, free/libre Radeon drivers still require proprietary firmware to run (3D acceleration, DPM, video decoding).

[Nille]