Bug 199655

Summary:	amdgpu: XFX Radeon RX 580 runs its fans only in dangerously low speeds and ignores temperature
Product:	Drivers	Reporter:	Sergey Kondakov (virtuousfox)
Component:	Video(DRI - non Intel)	Assignee:	drivers_video-dri
Status:	NEW ---
Severity:	high	CC:	0xe2.0x9a.0x9b, v.s.panasyuk, xomachiner
Priority:	P1
Hardware:	x86-64
OS:	Linux
Kernel Version:	4.16.7	Subsystem:
Regression:	No	Bisected commit-id:
Attachments:	dmesg_2018-05-08-quircks lspci_2018-05-08-quircks

Description Sergey Kondakov 2018-05-08 13:29:36 UTC

Created attachment 275835 [details]
dmesg_2018-05-08-quircks

In Windows AMD drivers like to ignore VGA BIOS fan control settings along with their own "Wattman" and disable fans until core temperature starts to go near 50-60 degrees, even at full load fans don't go over 2600 RPM there BUT on Linux they get stuck at 800-900 RPM by default (if pwm1_enable is not tempered with), at 1300-1400 RPM if pwm1_enable is set to '2', at 3500 RPM if it set to '0' and only manual control of '1' works as expected. temp1_* settings are outright ignored with error "permission denied". Unless manual control is used, GPU's core may overheat to >70 degrees (I don't even want to know what's happening on VRMs) at >90% load. This is madness.

In Linux on idle (~0% load in radeontop) GPU doesn't go lower than 41 degrees even with 1300-1400 RPM on fans, even though under Windows it goes to 35-40 with fans completely off (which I don't want to allow anyway because I don't know how safe are VRMs).

Easy way to overheat it is to use 'FSRCNNX_x2_r1_16-0-4-1.glsl' from https://github.com/igv/FSRCNN-TensorFlow/releases with a ≤720p video.
https://bugs.freedesktop.org/show_bug.cgi?id=103401#c2 - my modded BIOS with more aggressive cooling and lower frequency than stock.

Comment 1 Sergey Kondakov 2018-05-08 13:30:13 UTC

Created attachment 275837 [details]
lspci_2018-05-08-quircks

Comment 2 Sergey Kondakov 2018-07-03 07:07:49 UTC

Same thing with 4.17 kernel. It seems that any kind of fan speed control is happening only when pwm1_enable were completely untouched and there is high GPU load. However, it never changes memory frequency, so on idle it reports 40-44W of power usage instead of 12-15W on Windows. Now, in the middle of hellish summer, GPU temperature quickly rises from cold-boot 35 to 50-55 and stays there because of all that. Even my crappy FX-6100 CPU with 9mm fan isn't that hot on idle.

Comment 3 Uladzimir Panasiuk 2019-01-28 12:55:36 UTC

Same bug with 4.19.16 and Strix RX 470.