Nouveau Kepler On Linux 4.6 vs. NVIDIA's 364 Proprietary Driver

[phoronix]

Phoronix: Nouveau Kepler On Linux 4.6 vs. NVIDIA's 364 Proprietary Driver

We've already looked at the GeForce 900 series performance on Nouveau with the Linux 4.6 kernel that finally has the necessary bits for supporting 3D hardware acceleration on these Maxwell GPUs. Those GTX 900 series results didn't end up being particularly exciting since there isn't yet any re-clocking support on this open-source NVIDIA Linux driver. For complementing those results are some GeForce 600/700 "Kepler" graphics card results when comparing Nouveau on Linux 4.6 with Mesa 11.3-devel and using re-clocking then compared to the latest NVIDIA proprietary driver.

http://www.phoronix.com/vr.php?view=23014

[SpyroRyder]

Something I've always wondered in relation to Nouveau and the OS Amd drivers is how much of that is due to the drivers not being as good as opposed to the implementation not being as optimized. I suppose we'll get a good answer to this when AMDGPU gets to its fully usable state, but for the 680 in particular I wonder if a more optimized implementation would bump it up to the proprietary drivers level.

[theriddick]

They really need to take the magic from the 680 and put it in all the other hardware code. You would think it be easy to do.

[Vyrlokar]

Michael, do you have any Fermi (4XX/5XX) cards? I know they are old, but they're still in use.

[RSpliet]

@Michael: It's unfortunate that I couldn't find the system logs (dmesg) or output of the pstate file in OpenBenchmarking. I have the distinct feeling that the reasons for not achieving good performance on the 760 and 780 cards are due to incomplete voltage-setting logic in the 4.6 kernel, resulting in the core frequency not being touched when picking the 0f performance level for some but not all cards. Rather only memory is clocked higher.
If this is the case, there's some work currently being tested and in the process of being cleaned up (Karolherbst's tree) that will solve this issue. Unreleased because we need to make sure nouveau never sets the voltage lower than the official driver to avoid instability. System logs and/or pstate file output would confirm this theory. I thought pts automatically uploaded these logs; did I simply miss the link to the right location?

[Michael]

@Michael: It's unfortunate that I couldn't find the system logs (dmesg) or output of the pstate file in OpenBenchmarking. I have the distinct feeling that the reasons for not achieving good performance on the 760 and 780 cards are due to incomplete voltage-setting logic in the 4.6 kernel, resulting in the core frequency not being touched when picking the 0f performance level for some but not all cards. Rather only memory is clocked higher.
If this is the case, there's some work currently being tested and in the process of being cleaned up (Karolherbst's tree) that will solve this issue. Unreleased because we need to make sure nouveau never sets the voltage lower than the official driver to avoid instability. System logs and/or pstate file output would confirm this theory. I thought pts automatically uploaded these logs; did I simply miss the link to the right location?

Looks like the system logs didn't get uploaded for it. (There's a bug/safeguard where sometimes if the attempted result logs to upload are too large, they will be auto-rejected... Need to come up with better logic to decide when large logs can be safely kept.) Anyhow, yes, I intend to try out karolherbst 's tree with the latest code, hopefully later this week.

[andrei_me]

. I have the distinct feeling that the reasons for not achieving good performance on the 760 and 780 cards are due to incomplete voltage-setting logic in the 4.6 kernel, resulting in the core frequency not being touched when picking the 0f performance level for some but not all cards. Rather only memory is clocked higher.

Probably this answer the question of why only the 680 is achieving a good perfomance. Before your explanation I thought that the nouveau developers had more access/time to the 680 cards than the others...

[RSpliet]

Looks like the system logs didn't get uploaded for it. (There's a bug/safeguard where sometimes if the attempted result logs to upload are too large, they will be auto-rejected... Need to come up with better logic to decide when large logs can be safely kept.)

Oh that's too bad. I was going to ask if you had them dangling around, and if so whether you could e-mail the dmesg to me, but...

Anyhow, yes, I intend to try out karolherbst 's tree with the latest code, hopefully later this week.

if you have this planned anyway, that will probably answer my question too. No need to bother then. Thanks! :-)

[birdie]

Something quite weird is going on with GTX 760, because it's basically a slightly improved GTX 670, i.e. nouveau should yield very similar gains on GTX 680 and GTX 760, yet it doesn't.

[briantech]

Can reclocking be enabled by default when my system boots? If so how do I do it? I am not interested in a FOSS solution that requires manual intervention every time I run a game. But I have a 650, so these gains are encouraging. I would love to be able to run without blobs. I rarely render 3D anyway.