AMDGPU To Allow Memory Re-Clocking Soon For Multi-Monitor Setups

[perpetually high]

Many thanks to Alex Deucher! (and thank you for the heads up on this, Michael). I've had this problem with my RX 480 and two 144hz FreeSync monitors.

I just compiled the patches with 5.3-rc5 (the 5.2.x kernel didn't patch cleanly) and everything is working!

Using /sys/kernel/debug/dri/0/amdgpu_pm_info, here is the before and after on my RX 480.

Before:

GFX Clocks and Power:
2000 MHz (MCLK)
300 MHz (SCLK)
300 MHz (PSTATE_SCLK)
300 MHz (PSTATE_MCLK)
975 mV (VDDGFX)
25.37 W (average GPU)

After:

GFX Clocks and Power:
300 MHz (MCLK)
306 MHz (SCLK)
300 MHz (PSTATE_SCLK)
300 MHz (PSTATE_MCLK)
800 mV (VDDGFX)
10.109 W (average GPU)

AND, I just tried running WattmanGTK (adding amdgpu.ppfeaturemask=0xffffffff to GRUB like it said) and voila! WattmanGTK finally works and no more flickering. Really happy.

Verified FreeSync is still working with both monitors as well (I like to mirror the screens before gaming). WIll continue testing but so far so good!

[pal666]

will it work with 60hz monitor and 240hz freesync monitor?

[agd5f]

will it work with 60hz monitor and 240hz freesync monitor?

no. the monitors need the exact same timing.

[pal666]

no. the monitors need the exact same timing.

i understand this is current implementation state. but i imagine such monitors still could have simultaneous vblank 60 times per second like two 60hz monitors, that would be nice

[agd5f]

i understand this is current implementation state. but i imagine such monitors still could have simultaneous vblank 60 times per second like two 60hz monitors, that would be nice

This only works if the vblank periods are synchronized. The mclk switching has to happen during the vblank period or you get flickering when the clock switches. The timing has to be identical. Even if you have two monitors with the same refresh rates (60Hz, 75Hz, 120Hz, etc.), if they have different timings, the vblank periods won't not be aligned.

[pal666]

if they have different timings, the vblank periods won't not be aligned.

can vrr(freesync) be used to align them?

[agd5f]

can vrr(freesync) be used to align them?

The synchronization hw we use to synchronize the displays is designed to synchronize identical timings so it wouldn't help in this case.

[mczak]

Too bad the hw can't afford decent sized line buffers. Or alternatively just use L2 for scanout when it intends to switch mem clocks...

[perpetually high]

I can finally overclock my RX 480 and dual-monitor setup with this patch.

Went from a previously underclocked 1235 on the core and 2000 on the memory to 1290 (1100 mV) and 2140 (1000 mV). Anything at or over 1303 Mhz on the core (which is what MSI ships this card at!) crashes the system under full load (Rise of the Tomb Raider benchmark for example).

The same thing happens in Windows as well so it's likely the card. Had to underclock to have a working setup before today. In hindsight, i should have probably RMA'd the card a while ago.



This patch is what the doctor ordered, folks. Thanks again to those that made it happen.


EDIT: Did some more stress testing to figure out a stable overclock. 1290 on the core was no good - F1 2018 crashed after about an hour. Had to dial it back on the core to 1265 Mhz (1075 mV), but was able to bump memory to 2200 Mhz (also 1075 mV). All appears stable finally. My gpu is the MSI GAMING X RADEON RX 480 8GB in case anyone has similar stability issues and wanted to try out these clocks (of course, ymmv). Will update the post if anything changes.

[betam4x]

Windows user have been complaining a lot about this, they will be envious!

Maybe AMD users will, but my NVIDIA card downclocks the memory just fine...with 2 Freesync 4k displays no less.