You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The maximum memory clock for my Sapphire 7900 XTX Pulse according the official specs at https://www.techpowerup.com/gpu-specs/sapphire-pulse-rx-7900-xtx.b9966 are reported at 2500MHz, however using the profile-peak dpm level radeon-profile reports my memory clock being no more than 1249MHz, and I've never seen it go above that no matter what load, profile, level or mode I throw at it.
Using the auto dpm level (echo auto > /sys/class/drm/card0/device/power_dpm_force_performance_level ) and the compute powerprofile my core clocks easily go above 3.1GHz during intense ML workloads, but at the same time the memory clocks compensate by underclocking themselves to 96MHz, which has lead me questioning the memory clock values.
Here (youtube) Jayz overclocks a Powercolor 7900 XTX with the release Windows drivers and discusses the automatic memory underclocking, all the while the memory clocks aren't seen going below 909MHz during the whole video.
The level one sounds like the level jayz was getting underclock to on the above video.
Why do the amdgpu drivers report the memory clocks at half speed, is there some kind of "linux" rationale for this? According to this issue from the official AMD gitlab issue tracker https://gitlab.freedesktop.org/drm/amd/-/issues/1403 the MCLK was being reported at half-speed already with RDNA2. Should radeon-profile double the value or should I open an issue to the amd gitlab issue tracker?
The text was updated successfully, but these errors were encountered:
Foolish me didn't read the gitlab issue completely I went and linked in the first place, would be nice though if the GUI had an option to display all the MCLK values at double-data rate to avoid future confusion.
ojsl1
changed the title
Memory clocks being reported at half the actual value on 7900 XTX
Add an option to display memory clock values at double-data rate
Jun 3, 2023
The maximum memory clock for my Sapphire 7900 XTX Pulse according the official specs at https://www.techpowerup.com/gpu-specs/sapphire-pulse-rx-7900-xtx.b9966 are reported at 2500MHz, however using the profile-peak dpm level radeon-profile reports my memory clock being no more than 1249MHz, and I've never seen it go above that no matter what load, profile, level or mode I throw at it.
Using the auto dpm level (
echo auto > /sys/class/drm/card0/device/power_dpm_force_performance_level
) and the compute powerprofile my core clocks easily go above 3.1GHz during intense ML workloads, but at the same time the memory clocks compensate by underclocking themselves to 96MHz, which has lead me questioning the memory clock values.Here (youtube) Jayz overclocks a Powercolor 7900 XTX with the release Windows drivers and discusses the automatic memory underclocking, all the while the memory clocks aren't seen going below 909MHz during the whole video.
These are my sysfs memory clock levels:
The level one sounds like the level jayz was getting underclock to on the above video.
Why do the amdgpu drivers report the memory clocks at half speed, is there some kind of "linux" rationale for this? According to this issue from the official AMD gitlab issue tracker https://gitlab.freedesktop.org/drm/amd/-/issues/1403 the MCLK was being reported at half-speed already with RDNA2. Should radeon-profile double the value or should I open an issue to the amd gitlab issue tracker?
The text was updated successfully, but these errors were encountered: