Company of Heroes 2 Is The Latest Linux Game Showcasing AMD's Performance Wreck With Catalyst

[duby229]

Then why Nouveau (Gallium) and Intel (Mesa), smashes r600 and RadeonSI (Gallium), with smaller and cheaper hardware?

I think it's time for a Nouveau vs r600 vs radeonsi vs intel OSS driver comparison using the latest code.... It's already been done on this site so I'm sure some digging could find a link.

EDIT: Not exactly recent, or the best games to compare, but proves my point.
http://www.phoronix.com/scan.php?pag...ntel_316&num=1

[dungeon]

Then why Nouveau (Gallium) and Intel (Mesa), smashes r600 and RadeonSI (Gallium), with smaller and cheaper hardware?

Numbers seems fine, but are you sure rendering is fine there for 650?

Because I am not and why? Because Michael results are often broken - all Bioshock results are broken with Catalyst 15.7 it missrender if one does not apply "--eon_disable_catalyst_workarounds" with that render is fixed, game should be also around 15% faster.

Earlier with mesa Tesseract benchmark with radeonsi had broken render, numbers looks fine but he does not mentioned how it is borked... and so on... If something missrender i don't believe what benchmark numbers said - it is simple as that

CS:GO (and other source games) also depends on various in driver specific fixes, for performance it also depends on in driver threaded GL optimizations... which nvidia and flgrx drivers has implemented, but mesa not.

[dungeon]

And to not forget Dirt benchmark that one is particulary funny , because it runs in DX11 mode on Nvidia but on anything else in DX10 mode - so what we compare here? If you apply "--eon_disable_catalyst_workarounds" than it goes into DX11 mode. Same is with mesa, it runs with DX10 mode...

One may try "--eon_force_dx10" on nvidia for the same results

What all that means is when and because i know that, all numbers i see there tells me nothing... particulary no that Catalyst needs to be fixed, it might currently have couple minor flaws but that is about it

[smitty3268]

Then why Nouveau (Gallium) and Intel (Mesa), smashes r600 and RadeonSI (Gallium), with smaller and cheaper hardware?

First off, NVidia's driver not being especially faithful to the spec is a fact, but nobody said that's why it was faster. It's just why things tend to break on other drivers. (Though there may be some perf implications as well)

Second, although it's been awhile since i've seen good nouveau tests, you are very wrong about Intel being faster and cheaper than r600 or radeonsi. The hardware assures AMD of winning the vast majority of those tests over Intel, quite easily, and the only real exceptions are when you're taking $500 intel Iris Pro graphics and comparing them to $50 or $100 APUs from AMD.

It would be nice to see some Nouveau tests, though. Last time I checked, most of the hardware isn't capable of being clocked up without breaking things, unfortunately.

[bridgman]

Then why Nouveau (Gallium) and Intel (Mesa), smashes r600 and RadeonSI (Gallium), with smaller and cheaper hardware?

We seem to be talking about two different things. Everyone else is talking about non-compliant applications and non-standard usage of API calls, but AFAICS you keep jumping back to generic performance.

Not sure what "smashing... with smaller and cheaper hardware" you are talking about though - the only obviously faster open source performance I remember seeing is Iris Pro at >2x the price of comparable AMD HW. Can you be a bit more specific and a bit less dramatic ?

[artivision]

We seem to be talking about two different things. Everyone else is talking about non-compliant applications and non-standard usage of API calls, but AFAICS you keep jumping back to generic performance.

Not sure what "smashing" you're talking about though - the only obviously faster open source performance I have seen is on an Iris Pro at >2x the price of comparable AMD HW. Can you be a bit more specific and a bit less dramatic ?

I mean that a GF650 with Nouveau and full reclocking, on recent OpenGL games, gives 80%+ of the Nvidia driver performance and 2x vs your best GPU. What i mean is that recent OpenGL games they work good with the OGL Gallium state_tracker (as Nouveau shows), they are not only tuned for Nvidia driver. Also specific driver tuning doesn't exist to that level you say and doesn't produce those big differences, but different HW does.

[vadimg]

Define non-compliant apps. I you say that OGL games use extensions that only Nvidia has, then you are wrong, there are not such extensions. If you mean the way that each driver accepts them, then you speak for something that doesn't exist. SM4-5 are not prefixed but fully programmable, there are not non-compliant SM4-5 apps. So any compute shader can fail for example in a non-capable hardware. Nouveau (Gallium) has almost the performance of Nvidia-driver when reclocked. Its not the drivers or the game, if there was they would have fix it. I propose to AMD to start with new MIMD hardware and smaller units.

Sorry, but you really seem to have no idea what are you talking about. Did you ever write any OpenGL app, maybe just a small demo? You say there are no NV-only OpenGL extensions, really? Moreover, there are many extensions supported by both NV and AMD, but they just work worse on AMD due to hardware differences. So on AMD for optimal performance you need other extensions, etc. Anyone who tried to develop something that works on both AMD and NV knows it.

Not even sure what you are saying about compute, AMD hardware easily beats NV there, even Michael's benchmarks confirm it.

I just hate it when e.g. blender developers write some ugly code and blame AMD drivers for not being able to handle it. First you should understand how to use GPU for acceleration. Then write a good code, and you can acheieve acceleration. Otherwise you are wasting time. IIRC AMD devs had to finally rewrite blender stuff, I'd be very ashamed of it if I were blender dev, just means you don't even understand what are you trying to do.

[bridgman]

I mean that a GF650 with Nouveau and full reclocking, on recent OpenGL games, gives 80%+ of the Nvidia driver performance and 2x vs your best GPU.

I can't find any benchmark results that support this, can you be a bit more specific please ?

Best results for GTX 650 with Nouveau at 1920x1080 seem to be ~52 fps while tests from the same time period at 2x the resolution suggest comparable performance (lower fps but more pixels) and more recent tests at same resolution with different AMD GPUs show 52-96 fps.

http://www.phoronix.com/scan.php?pag...318_blob&num=2

http://www.phoronix.com/scan.php?pag...on_12nov&num=1

http://www.phoronix.com/scan.php?pag...on_12nov&num=1

We don't really have a precise equivalent to GTX 650 (low shader count but high ROP & memory bandwidth) but 7790 / 260x seems to be closest.

[dungeon]

Bridgman you have Kaveri around that fps in CS:GO

62 fps on Windows:

http://hexus.net/tech/news/cpu/83423...-unlocked-apu/

58 fps on Linux:

http://www.phoronix.com/scan.php?pag...ri-steam-linux

With cpufreq ondemand With right profile it should be fired up above 80 fps probably

edit: blah, if that happens it would kill all nouveau running cards

[pete910]

Thought some of you may find this interesting.

http://www.overclock.net/t/1569897/various-ashes-of-the-singularity-dx12-benchmarks/1200#post_24356995

Difference is this Dev didn't do what NV wanted. Good for them I say.