Nouveau Linux 4.8 + Mesa 12.1-dev vs. NVIDIA OpenGL Performance

[dungeon]

wrong though. In this case those GPUs aren't fully reclocked. The GTX 680 is fine though.

Ah OK, i tought Ehvis asked for remaining gap even on that GTX 680 regardless it is reclocked.

[karolherbst]

Shorter explination might be... that something is bound you know and that need to be optimized

wrong though. In this case those GPUs aren't fully reclocked. The GTX 680 is fine though.

[karolherbst]

well, with "* adjusting clocks on temperature changes" I more mean like changing the max clock. Basically with GPU Boost you have a voltage formula attached to each clock state (there are up to 60 of those) and depending on the temperature the resulting voltage differs, same goes for the maximum voltage.

Usually the max voltage entry are defined in a way, that the voltage drops the higher the temperature, which means some cstates might become unavailable out of the sudden and we have to dynamically adjust to this.

[karolherbst]

What I don't understand is how this reclocking factors into all this. If I understand correctly, then the "0F" setting means forcing the highest clock. In theory this means that it should have the highest performance. Yet, if this is the case, then there is still a massive gap to the proprietary driver that has nothing to do with reclocking.

Is there a proper explanation for that?

basically without my patches reclocking is broken, it just works on some cards by pure luck. It isn't broken in a sense that it would crash the card (it might though), but the entire GPU Boost thing wasn't implemented and there is a lot of things to be aware of. Even with the landed patches there is still a long way:

* adjusting clocks on temperature changes
* power capping

the former is pretty much done and currently in review, but the latter will still take some time.

[dungeon]

Is there a proper explanation for that?

Shorter explination might be... that something is bound you know and that need to be optimized

[Michael]

What I don't understand is how this reclocking factors into all this. If I understand correctly, then the "0F" setting means forcing the highest clock. In theory this means that it should have the highest performance. Yet, if this is the case, then there is still a massive gap to the proprietary driver that has nothing to do with reclocking.

Is there a proper explanation for that?

karolherbst can probably explain the best, but without the boost patches (now for Linux 4.10) the 'boost' clock frequencies aren't reached for newer cards. I think there also might be cases where vRAM re-clocking isn't done or not to the highest state, but Karol can probably say whether that's still an issue or not.

[Ehvis]

What I don't understand is how this reclocking factors into all this. If I understand correctly, then the "0F" setting means forcing the highest clock. In theory this means that it should have the highest performance. Yet, if this is the case, then there is still a massive gap to the proprietary driver that has nothing to do with reclocking.

Is there a proper explanation for that?

[peppercats]

Wish AMD cards were competitive with nvidia on proprietary driver. The highend AMD cards still run like ass.

[cRaZy-bisCuiT]

it is a little sad that the nouveau updates weren't ready to be pushed to drm-next and won't land in 4.9 :/ There are some ideas though to make it much easier for users to patch a stock kernel and get updated drm/nouveau to play with. Maybe we get something ready soon.

Sounds good! Still, considering the benchmarks, it seems like a long way to go. With other words: There's still a lot of potencial!

[dungeon]

It is not nothing, at least once card result looks usable.