Here's Why Radeon Graphics Are Faster On Linux 3.12

[log0]

Additionally, the game locks threads on cores.

Thanks for the link. So the Croteam guys have figured out the reason quite a while ago...

We have determined through testing that the "ondemand" governor (which is default, and which you have) seems to have some bugs and causes the erratic behavior that you see on your FPS graph. It appears as if the governor and scheduler interact in a weird way so that governor downclocks less used cores, but then the scheduler moves the main thread to that underclocked core, which the governor then clocks back, etc ad nauseam. We have contacted kernel devs, most notably developers from Intel that are working on that area, but they seem to still be looking into this.

The only workaround that we can come up from application side was to set strict affinity for each thread.

Ultimately, it is best if you switch to using the "performance" governor.

[Kariddi]

So basically this confirms that benchmarking with games that output 300 FPS is completely a waste of time ....

[JS987]

Intel CPUs are supported also by newer P-state driver with performance governor which should scale frequency like old ondemand one, but more reliably.

[wargames]

Does this affect all CPUs or only some specific ones? I own a Phenom II x6.

[marek]

Why are we using ondemand for benchmarking in the first place? Benchmarking should be done to see what the system can do in terms of performance and using performance governor should be the "default" in that scenario. Does Michael use ondemand because he's interested in power consumption as well?

I second this. People who are serious about performance should always set the "performance" governor. I even removed the file /etc/init.d/ondemand, so that I always get "performance" on boot. The init.d script sets "ondemand" after 1 minute or so after boot, i.e. when you least expect it, which messes up benchmark results. If you play games with the "ondemand" governor and complain about performance to us, you're wasting our time.

While it's a good thing that we know what the problem was, it's also a big fail for Phoronix. "ondemand" should never ever be used for benchmarking.

[Figueiredo]

So, in a nutshell, an intel core i7 4770k can still bottleneck the R600g driver.

Michael, since more demanding games that are capable of stressing the GPU rather than the CPU with fully automated bechmarking are not available, could you include CPU utilization monitor into PTS? This seems to be a really import variable, especially to gallium drivers since intel's mesa driver does not seem too suffer from the same bottleneck.

[89c51]

While it's a good thing that we know what the problem was, it's also a big fail for Phoronix. "ondemand" should never ever be used for benchmarking.

Average joe configuration should be used IMO. Also the user MUST NOT have to care about stuff like that. It should just work.

[pheldens]

btw. that commit is dated july this is around the time the problems with ondemand were discovered and already discussed on phoronix.

So are these tweaks ontop of it still needed?

echo "tuning CPU ondemand governor..."
echo "15" > /sys/devices/system/cpu/cpufreq/ondemand/up_threshold # 95
echo "10" > /sys/devices/system/cpu/cpufreq/ondemand/sampling_down_factor # 1

[jsa1983]

I second this. People who are serious about performance should always set the "performance" governor. I even removed the file /etc/init.d/ondemand, so that I always get "performance" on boot. The init.d script sets "ondemand" after 1 minute or so after boot, i.e. when you least expect it, which messes up benchmark results. If you play games with the "ondemand" governor and complain about performance to us, you're wasting our time.

While it's a good thing that we know what the problem was, it's also a big fail for Phoronix. "ondemand" should never ever be used for benchmarking.

You could call it a usability test rather than a benchmark In any case, I think it's important to get the default settings working OK for the most people so I welcome any improvement to the ondemand governor.

[AnonymousCoward]

It just means the open source AMD radeon driver depends too much on the cpu instead of, duh, the processing power of the graphic card.
The other drivers are not affected because they actually use the graphic card instead of the CPU.

Nouveau results are still outstanding, but seeing the faster Radeons being affected the most by this, Intel's GPUs might just be too slow to need much CPU to keep them fed with data to render.