An Interesting Difference Between AMD & NVIDIA Linux Drivers When Comparing System Usage

[SystemCrasher]

Similarly, the GPU frequencies / performance states seemed to bounce around a lot more on the AMD

Once again, there is no "GPU frequency" on AMD these days. GPU reclocks itself like 100 times a second to match actual load. So measuring it in terms of fixed frequency is rather pointless. At most something like average frequency can make sense, but if you claim it "bounced", you're likely measured something else. So it raises question about what exactly has been measured, granted how AMD reclocking works.

[kruger]

Wasn't it already proven many times that AMD driver is often CPU bound? It would make sense that GPU clocks itself down when it's not getting data/commands on time.

[bug77]

Wasn't it already proven many times that AMD driver is often CPU bound? It would make sense that GPU clocks itself down when it's not getting data/commands on time.

That's not the only find in the article. It seems the system memory usage is also higher.

[zuxun]

The AMD gpus seem to have a very low average frequency.

I wonder how this affects the performance.

[GreatEmerald]

Hm, interesting indeed. I'm not sure I'm convinced by the "different code paths" argument here in the comments; why would they be different, for one, and even if the code paths are different and one of them (the AMD one) results in higher RAM usage, that's still higher RAM usage in the end. I don't think games are coded to use more RAM on one platform and less on another, if the platform itself doesn't do something to eat more RAM. On the other hand, simply the AMD blob being bigger than the NVIDIA blob is quite possible.

As for the frequencies... That looks bad on the AMD side to me. If it was CPU-bound, then... you would get more CPU usage using the AMD blob, which is clearly not the case. Mind you, it could still be CPU-bound *and* not using the CPU because of some problems in the blob, but that's still a problem in the blob; whether it's in the frequency algorithm or CPU utilisation algorithm isn't that important in the end.

It would be nice to see the same with the free drivers. Generally it makes me hopeful that radeonsi can in fact overtake fglrx eventually.

[bug77]

Hm, interesting indeed. I'm not sure I'm convinced by the "different code paths" argument here in the comments...

It's a red herring. The cause is of less importance, the article is about the effects.

[profoundWHALE]

Just looking at something here, I took the 290 CPU bench and divided by the 780 Ti CPU bench. All numbers rounded to the 2nd decimal and using the Metro specific benchmarks where possible.
R290 80.08% MAX CPU Usage vs 780 Ti
R290 93.53% AVG CPU Usage vs 780 Ti

Now the same with GPU Frequency, normalized by percentage (X / 947 = R290 % Frequency, X / 1019 = 780 Ti % Frequency)
R290 GPU AVG 88.70% MAX Frequency
R290 GPU MIN 31.68% MAX Frequency
780 Ti GPU AVG 94.50% MAX Frequency
780 Ti GPU MIN 31.80% MAX Frequency

R290 93.86% AVG GPU Frequency vs 780 Ti
R290 99.62% MIN GPU Frequency vs 780 Ti

Now the same with GPU Usage
R290 92.26% AVG GPU Usage vs 780 Ti

CPU Usage * GPU Usage / 100 =
CPU vs 780 Usage * GPU vs 780 Usage / 100 = 86.29% CPU and GPU usage

Now the same with FPS for Metro
R290 70.91% AVG FPS vs 780 Ti
R290 AVG GPU Usage * R290 vs 780 Ti AVG CPU Usage / 100 = 71.03%

Any possible correlations?

[waxhead]

I don't usually pay much attention to graphics card benchmarks anymore , but isn't the only useful metric here the average cpu usage pr frame produced by the graphics card?! (Of course the average memory used per frame is also interesting to compare with the previous one).

[Stellarwind]

Agree with waxhead. CPU usage should be measured per frame, unless you do some synthetic benchmark, just rendering same scene over and over again. Otherwise it is comparing oranges to apples.

[theriddick]

What kernel do you use for 15.10? is it 4.2? With that Kernel my 390x performs terribly with the latest crimson drivers, to the point of me quitting Linux for the time being. They also seem to crash allot, for example atm I have not touched my Linux install but when I boot it up the display driver crashes the system. I'll get around to fixing it someday but it's real annoying to get bad performance and crashing with crimson and kernel 4.2