amdpwrman is now being replaced by a newer app, gpumgr. Currently, gpumgr only implements the show command from amdpwrman but will soon support all it's features, have a properly functioning fan curve daemon, a web interface and ultimately add support for different GPU vendors like nvidia.
Feel free to use amdpwrman in the meantime but definitely check out the new project as that's where my efforts will be going.
amdpwrman 0.05a (C) 2020 Jai B. (Shaped Technologies) MIT License
amdpwrman shows statistics and manipulates power limit settings for AMD GPUs on Linux through the sysfs interface provided by the amdgpu driver. This script was designed to be very simple to use from the shell, requiring no external dependencies or configuration, though a future version may be based on NodeJS or be a binary application - the goal will still be to be as light and easy as possbile.
All values are reported by the sysfs interface provided by the amdgpu driver, sometimes these values may not be perfect, for example on an RX570, the max RPM of the fan is reported as 5300 RPM but at 100% fan speed the fan only hits 3300 RPM. I've also noticed this behavior with apps on Windows even, so it is not isolated to Linux, nor is it an issue with this script.
Note: amdpwrman will require root access to modify values. Attempting to recover a GPU may require the amdgpu.gpu_recovery=1 kernel flag.
This script was created because I needed a simple way to manage my AMD mining GPUs from the console.
Usage:
If <gpu> is ommitted from any command, GPU0 is used.
amdpwrman help | -h Display this help message.
amdpwrman show <gpu> Shows detailed statistics for <gpu>
amdpwrman status <gpu> Same as above
amdpwrman power <limit> <gpu> Sets power limit to <limit> in watts for <gpu>
amdpwrman power reset <gpu> Resets power limit for <gpu>
amdpwrman recover <gpu> Attempt to recover a halted or crashed <gpu>
amdpwrman fan enable <gpu> Enables manual fan speed control for <gpu>
amdpwrman fan disable <gpu> Disables manual fan speed control for <gpu>
amdpwrman fan <speed> <gpu> Sets fan speed for <gpu>
amdpwrman fancurve enable <gpu> Enables fan curve daemon for <gpu>
amdpwrman fancurve disable <gpu> Disables the fancurve daemon for <gpu>
<gpu> refers to the number of the GPU reported by the amdgpu driver.
The fan curve daemon does not need to be running for manual fan control,
it only needs to run if you want to define custom fan curves, in which
case the daemon will read from ~/.amdpwrfan to determine the curves,
which can be set by manually editing the file or using the fancurve set
command, which will create the file for you. See README.md for an
example config, or simply 'use amdpwrman fancurve enable' without a
config and amdpwrman will create a default one for you.
[fancurves] should be formatted as follows:
temp:speed for every temperature/speed you want to add
values between specified values will be interpolated
Example: To set the fan curve to 0% at 25 C, 25% at 40 C, 50% at 50 C
and 75% at 60 C, updating every 5 seconds for GPU 0, you would use:
sudo amdpwrman fancurve set 5 25:0 40:25 50:50 60:75 65:100 0
Examples:
#$ ./amdpwrman show Show stats for GPU0 (TODO: show stats for all gpus)
#$ ./amdpwrman power 100 Sets power limit to 100W for GPU0
#$ ./amdpwrman power 100 3 Sets power limit to 100W for GPU3
#$ ./amdpwrman fan enable 3 Enables manual fan control of GPU3
#$ ./amdpwrman fan 100 3 Sets GPU3's fan speed to 100%
#$ ./amdpwrman fan disable Disables manual fan control of GPU0
#$ ./amdpwrman power reset Resets the power limit of GPU0
FAQ:
1q. Why?
1a. Why not? I needed an easy way to do this and got sick of manipulating the files by hand, so I whipped up this bash script in a few hours.
2q. Why AMD?
2a. While historically I have generally prefered NVidia cards, I do use AMD cards for mining currently as they are more economical (An RX570 provides about the same hashrate as a RTX2060 for Etherium for example).
3q. Will you support NVidia?
3a. I'd like to, I haven't looked into what interfaces are available for the NVidia drivers. I currently don't have any suitable NVidia cards in any Linux boxen to test with. Maybe one day.
4q. Can I contribute?
4a. If you have additions that may be useful or find any bugs, feel free to send them in.
5q. MIT License?
5a. Yep. This is a quick hacky shell script that I spent more time on than I planned and figured it might be worth sharing. Do whatever the fack you want with it, just maybe at least refer to me as the original author though.
TODO:
- support multiple GPUs at once, eg. 'amdpwrman show all'.
- change 'show' without a GPU specifier to show all GPUs once multigpu support is complete.
- decide how to implement fan curve daemon; should we use node? If so, should the whole thing be node? If not, how bad is it really to run a shell script as a daemon?
- the no-dependencies nature of using shell script is great for this application. however, shell scripting isn't the cleanest way to code and definitely has limitations.
- many distros, especially LTS, ship a realy old version of node. not too hard to work around but still a pain.
- can we compile a sort of node binary? bytecode/opcodes? packed interpreter/vm? I suppose this could be an option with PHP bytecode/opcodes too, while php would normally be avoided for the same dependency reasons.
- fan curve control has started to be implemented
- fan curve control currently only selects the next highest fan setting and uses that
- fan curve control needs to interopolate fan speeds between settings instead of simply going to the next highest
- fan curve control needs to be tweaked to support multiple GPUs at once - while it's been structured for multilpe GPUs from the get-go, it currently only will allow one instance to run. simply allowing multiple instances would allow fan control of multiple GPUs but I feel a single instance for all GPUs is best
- fan curve control format is: delay temp:fanspeed ... temp:fanspeed gpu .. if we are managing all GPUs in one instance, then we need to either have the same delay for all GPUs, which means another way of setting it instead of it simply being the first variable in the list OR we need to actually do some more fun stuff so that it updates each GPU based off it's waittime
- fan curve control should have options for linear and log scaling/interop
CHANGELOG:
0.01a first release. can control fans and power and show some stats.
0.02a shows some more stats
0.03a shows even more stats (including current power profile and gpu/mem clks)
0.04a started implementing fancurve control. this is in alpha.
0.04b just a fix for multiple GPUs
0.05a implemented a fix for Issue #1
0.05b implmented fix for Issue #2 that arose from the half-assed fix to issue #1