Experimental Nouveau DRM Branch Yields Better GDDR5 Kepler Re-Clocking

[phoronix]

Phoronix: Experimental Nouveau DRM Branch Yields Better GDDR5 Kepler Re-Clocking

If you've been reading Phoronix any length of time for open-source graphics news, you'll know the most difficult challenge facing open-source NVIDIA (Nouveau) driver developers is the re-clocking / power management situation. It's been slow and re-clocking still leaves a lot to be desired...

http://www.phoronix.com/scan.php?page=news_item&px=Nouveau-GDDR5-Branch

[Pecisk]

Yes! Most importantly, please report back if *switching* between modes are stable (going to 0a -> 0f -> 0e -> 0f and so on and so forth). This is something developers are looking after. Also check if core frequency are at the floor of the mode - for example f0 did this for me - this is known bug due of voltage mismatch and there's one clever workaround with modifying cstate - I will try to add more information to my subreddit article.

[karolherbst]

@michael

there are still some issues left on kepler cards:

1. some cards get a wrong voltage sets at higher clock, so these cards will still crash at 0f pstate
2. the change to 0f is not rock solid yet either, so sometimes there are corruption on the screen, but changing back to 07 and 0f again may fix that.

Just two pointers while doing your tests

[kuco]

Why is reclocking so hard to implement? Can somebody explain?

[edmon]

Why is reclocking so hard to implement? Can somebody explain?

https://en.wikipedia.org/wiki/Black_box

[karolherbst]

Why is reclocking so hard to implement? Can somebody explain?

because you have to deal with a lot of stuf:
Timings, PLLs, voltage control, parsing the vbios, suspending the engine(s). ...

it's not a simple write into one or two regs, but more likely 50 of them, which you need to write and read, if you want to have it rock solid. And the pain is, nobody knows which are the regs you have to use, so you need to analyze what the blob is doing and so on.

[Pecisk]

Why is reclocking so hard to implement? Can somebody explain?

In this case because there's no specification for this. It's called black box reverse engineering and that's how whole Nouveau project exists, because Nvidia has never published driver/protocol specs for their mainstream cards.

So devs have to use clever tools to extract information bit by bit trough trial and error (also prior knowledge how stuff works in industry in general helps). Only then they can implement actual code, which is comparingly simple. Getting there is hardest part.

[Pecisk]

Why is reclocking so hard to implement? Can somebody explain?

Reclocking itself is relatively easy. Problem is there's no spec for Nvidia cards, so everything is subject of reverse engineering. It takes time to understand what is proper way to increase and decrease voltage.

[karolherbst]

Reclocking itself is relatively easy. Problem is there's no spec for Nvidia cards, so everything is subject of reverse engineering. It takes time to understand what is proper way to increase and decrease voltage.

allthough voltage is one of the smaller problems, understanding the vbios is a much bigger problem here, because if the vbios is fully understand, setting the right voltage is piece of cake

[kuco]

Reclocking itself is relatively easy. Problem is there's no spec for Nvidia cards, so everything is subject of reverse engineering. It takes time to understand what is proper way to increase and decrease voltage.

yeah i know. i was just wondering, because i remember playing with nvdia clockrates/voltages by modifying the vga-bios. i assumed, REing the bios (not very big) is easier than REing that giant blob of a driver.