Nouveau "Boost" Patches Show Much Performance Potential

[karolherbst]

I didn't mean the core clock.


I mean, the memory clocks.


If I try to relax it to, for instance, 03, after a 0f state, this is what I see :

03: core 300 MHz shader 600 MHz memory 100 MHz AC DC *
07: core 400 MHz shader 800 MHz memory 300 MHz
0f: core 576 MHz shader 1242 MHz memory 999 MHz
AC: core 300 MHz shader 601 MHz memory 999 MHz

And dmesg shows me "nouveau 0000:01:00.0: fb: Could not calculate MR" when I try a lower pstate.

Don't happen when I switch to a higher.



PS : GTX 260, Linux 4.4, 64 bit.

Mhh this is Tesla. I never looked into it, but I guess the issue is something trivial. You should open a bug if you didn't do it already. And if nobody commented on that, ask again

[DebianLinuxero]

I didn't mean the core clock.


I mean, the memory clocks.


If I try to relax it to, for instance, 03, after a 0f state, this is what I see :

03: core 300 MHz shader 600 MHz memory 100 MHz AC DC *
07: core 400 MHz shader 800 MHz memory 300 MHz
0f: core 576 MHz shader 1242 MHz memory 999 MHz
AC: core 300 MHz shader 601 MHz memory 999 MHz

And dmesg shows me "nouveau 0000:01:00.0: fb: Could not calculate MR" when I try a lower pstate.

Don't happen when I switch to a higher.



PS : GTX 260, Linux 4.4, 64 bit.

[karolherbst]

If that's so, then I concur that my card doesn't reclock correctly. It doesn't relax the memory speed in slow down transitions.

What do you mean?

At least the Core clock doesn't need to be equal to any of the clocks above. It is only guaranteed that it is equal or lower to the currently set boost level and equal or lower to the pstate clock.

[DebianLinuxero]

Ummh.

If I undertand well, the bottom line tagged "AC" shown when you cat pstate file, shows the real clocks at what the card is now isnt it?

So if all goes well, that line should be equal to one of the others.

If that's so, then I concur that my card doesn't reclock correctly. It doesn't relax the memory speed in slow down transitions.

Only speed it up on speed up transitions..

[karolherbst]

A question targeted at those which are closer to the nouveau development: is similar work being done for Fermi based cards? (I have a work notebook with a GF108GLM/5200M and nouveau cannot drive the card (fast enough?) so that two big external monitors can be used without artifacts - so I have to use the proprietary driver for now.)

Well in the end this boosting stuff is also interessting for many Fermi cards, but for Fermi Memory reclocking has to be implemented. In the end it just means, when Memory reclocking is done for Fermi, the entire Fermi reclocking situation will be pretty solid from the start.

[OlafLostViking]

A question targeted at those which are closer to the nouveau development: is similar work being done for Fermi based cards? (I have a work notebook with a GF108GLM/5200M and nouveau cannot drive the card (fast enough?) so that two big external monitors can be used without artifacts - so I have to use the proprietary driver for now.)

[Julius]

Interesting... if Noveau can give hassle free optimus support, I am sure quite a few people would accept a not that huge performance impact. Looking forward to 2017

[karolherbst]

Hello.

A question :

I'm using nouveau 2.4.67 on a GTX 260 over Debian Testing 64 bit (Kernel 4.4).

I activate the pstate option on grub2.

As I could see, the pstate defaults to 07 at every system startup.

Doesn't suppouse to default to 03 either?

No, it defaults to what the GPU sets, because touching it may break things.

[DebianLinuxero]

Hello.

A question :

I'm using nouveau 2.4.67 on a GTX 260 over Debian Testing 64 bit (Kernel 4.4).

I activate the pstate option on grub2.

As I could see, the pstate defaults to 07 at every system startup.

Doesn't suppouse to default to 03 either?

[karolherbst]

nope, GL_ARB_buffer_storage is just a big broken for nouveau.