Radeon Linux Benchmarks: Catalyst 15.3 Beta vs. Linux 4.0 + Mesa 10.6-devel

[peppercats]

4. Optimizing the driver for frame rates much over 100 Hz is mostly useful for benchmark competitions and the open source folks don't have to care about them. Catalyst is a different story because companies still make buying decisions based on one card running 275 FPS and another running 297 FPS, even though they're both being used with a 60 Hz display.

which is why FPS graphs should be logarithmic - the first 30 FPS are far more valuable than 100 extra FPS at 400(...which is actually only shaving .5 ms/frame off, 30->31 FPS(33.3 ms/frame -> 32.25 ms/frame) is a bigger jump than 400->500(2.5 ms/frame -> 2 ms/frame.))

if you convert these graphs to logarithmic you'll realize the difference between them at FPS that matter is actually pretty small :/

[xeekei]

4. Optimizing the driver for frame rates much over 100 Hz is mostly useful for benchmark competitions and the open source folks don't have to care about them. Catalyst is a different story because companies still make buying decisions based on one card running 275 FPS and another running 297 FPS, even though they're both being used with a 60 Hz display.

What about multiple monitors? Achieving high frame rates on something like 1080p, or even 1440p is one thing, but 6x1080p is another. I'd assume AMD keeps this in mind since you were pushing multimonitor with "Eyefinity".

The "this is enough" mentality is an anchor for innovation.

[smitty3268]

Not really seeing as AMD isn't using LLVM 3.7 for the Catalyst driver, until LLVM 3.7 is out.

They aren't using Mesa 10.6 in the Catalyst driver either....

Not sure what your point is.

[humbug]

Catching up with Catalyst is not as obvious an outcome as you might expect because performance work on Catalyst is still going on as well, so while the open source driver keeps getting faster the Catalyst stack keeps getting faster too.

Good to hear as an R9 290 user.

Has there been any progress in identifying/eliminating the bottleneck which causes framerates to hit a wall at a very low frame rate in newer openGL 4 games such as Metro Redux, witcher 2, Civilization BE etc?
From the outside looks like a CPU bottleneck where the GPU is not drawing significant power, it's not working hard to produce more frames. I have tested this in my system too and my GPU temperature stays very cool under gaming loads in Linux compared to windows.

[humbug]

which is why FPS graphs should be logarithmic - the first 30 FPS are far more valuable than 100 extra FPS at 400(...which is actually only shaving .5 ms/frame off, 30->31 FPS(33.3 ms/frame -> 32.25 ms/frame) is a bigger jump than 400->500(2.5 ms/frame -> 2 ms/frame.))

if you convert these graphs to logarithmic you'll realize the difference between them at FPS that matter is actually pretty small :/

Or... People should just be intelligent enough to interpret the data, LOL

[asdfblah]

Catching up with Catalyst is not as obvious an outcome as you might expect because performance work on Catalyst is still going on as well, so while the open source driver keeps getting faster the Catalyst stack keeps getting faster too. Main focus for the open source developers right now is adding GL features and new HW support, not performance tuning.

What about comparing r600/radeonsi against the closed-source nvidia driver in terms of performance (outside benchmarks...)? The catalyst vs nvidia benchmarks show terrible performance for the former...

[SXX⁣]

What about multiple monitors? Achieving high frame rates on something like 1080p, or even 1440p is one thing, but 6x1080p is another. I'd assume AMD keeps this in mind since you were pushing multimonitor with "Eyefinity".

If you talk about games then 400FPS on 1080p monitor doesn't mean that on 6x1080p you'll have 66FPS. This is because GPU performance not just limited to how fast computation parts are, but VRAM amount and bandwidth is main bottlenecks.

And for some post processing and things like anti-aliasing amount of VRAM may be more important than performance of GPU.

[peppercats]

Good to hear as an R9 290 user.

Has there been any progress in identifying/eliminating the bottleneck which causes framerates to hit a wall at a very low frame rate in newer openGL 4 games such as Metro Redux, witcher 2, Civilization BE etc?
From the outside looks like a CPU bottleneck where the GPU is not drawing significant power, it's not working hard to produce more frames. I have tested this in my system too and my GPU temperature stays very cool under gaming loads in Linux compared to windows.

the bottleneck doesn't happen in any open source game so it's likely due to shitty ports.
nvidia proprietary driver is heavily optimized to deal with shitty opengl practices, AMD not so much.

[Luke]

Even though RadeonSI has more to go, new cards still tested faster

An interesting part of the results is the graph of just the open source driver framerates: Those big RadeonSI cards in the test easily beat the older (though smaller) r600 cards. The fact that they beat them by even more in the closed driver test just means there is more performance yet to be found in tweaking radeonSI. Those cards will fly when radeonSI gets as good as r600 is right now.

Someone mentioned a regression on r600 with one of the very large cards, I have not had one on the Radeon HD6750 nor on the HD5570 since a year ago when there were SDL issues with a new version of X, long since resolved. I've been quite happy with the r600 driver on my hardware and do not have any games or other loads that it can't handle. Of course I am speaking as one who also does not play closed games. Somehow a closed came demanding closed drivers is about a surprising to me as a Windows program only running well (or at all) under Windows.

[dungeon]

the bottleneck doesn't happen in any open source game so it's likely due to shitty ports.

Not quite true. This just as example, even 15 years old games can hit slow cases For example default cvar 'seta r_primitives 0' in many quake 3 engine based games is 2.5 times slower with fglrx for quite some time now , you need to set it on 2 to make it fast, google about it and io code is there