Nouveau Kepler Re-Clocking Is Working A Lot Better On Linux 4.4

[Michael]

Michael,

as imirkin said, something is odd with the gpu core clock with your 780 ti. If you want you could give us access to your vbios.rom (located in /sys/kernel/debug/dri/0 as root) and we can have a look at this. There are some voltage realted issues pending on my side I fixed for two kepler cards, but this needs a little bit more investigation, so it might take some time.

On my side I have a 2.5x speed up form 07 to 0f and +33% from 0a to 0f, but my gpu core also reclocks.

Sure, can send over the vbios of the 780 Ti next time I have it plugged in.

[Michael]

Michael,

Looks like the core speed got stuck at the lowest level, which is why you're seeing only minor improvement with faster memory. Can you try clocking to other levels first to see if that helps? (A different level might have different cstate requirements which will allow the core speed to change.) I suspect you're seeing voltage-change-related errors in dmesg when doing this... it will take a while to iron out all of those board-specific issues.

Perhaps in the meanwhile you can find a board in your collection that doesn't hit this issue? Just look to make sure that the speeds on the "AC" line match up with the high end of whatever perf level you selected. (Ignore the * on the perf level line, it's just there to lull you into a sense of false safety.)

I have tried on the GTX 780 Ti to re-clock to the other levels first, but for that card at least it's still a no go. I have some other results I had already done that will be publishing shortly.

[Kano]

Then use the i5-6600K but no i7-5960X. Best ask Intel for i7-6700K.

[Michael]

Michael

Can you use an i7-6700K for benchmarks instead of a low clocked i7-5960X cpu?

I don't even have a 6700K.

[Kano]

Michael

Can you use an i7-6700K for benchmarks instead of a low clocked i7-5960X cpu?

[karolherbst]

Michael,

as imirkin said, something is odd with the gpu core clock with your 780 ti. If you want you could give us access to your vbios.rom (located in /sys/kernel/debug/dri/0 as root) and we can have a look at this. There are some voltage realted issues pending on my side I fixed for two kepler cards, but this needs a little bit more investigation, so it might take some time.

On my side I have a 2.5x speed up form 07 to 0f and +33% from 0a to 0f, but my gpu core also reclocks.

[pjezek]

Unfortunately, no strong strategy for Nvidia OSS driver's development continues keeping use of nouveau and Nvidia GPUs risky under Linux... Very crazy strategic management in Nvidia headquaters...

[imirkin]

Michael,

Looks like the core speed got stuck at the lowest level, which is why you're seeing only minor improvement with faster memory. Can you try clocking to other levels first to see if that helps? (A different level might have different cstate requirements which will allow the core speed to change.) I suspect you're seeing voltage-change-related errors in dmesg when doing this... it will take a while to iron out all of those board-specific issues.

Perhaps in the meanwhile you can find a board in your collection that doesn't hit this issue? Just look to make sure that the speeds on the "AC" line match up with the high end of whatever perf level you selected. (Ignore the * on the perf level line, it's just there to lull you into a sense of false safety.)

[jaxxed]

I see finally NV's Open-source drivers a future. That's great!

Sorry to be "that" guy but:

We're gonna need those signed firmwares to be available, so that the nouveau guys aren't blocked.

Otherwise, we're gonna have to assume that NVIDIA is trying to kill the "open" part of it, and push it to being a Tegra only thing.

[adakite]

This is great news....In addition to the compute capabilities that will land one day or another with Nouveau along with pure graphical usage, I see finally NV's Open-source drivers a future. That's great!