Linux 5.20 To Have Power Management / Turbo Fix For AMX-Enabled Intel CPUs
Last month I wrote about a new set of Linux patches that provide for better power management on AMX-enabled "Sapphire Rapids" servers. The patches help ensure AMX-enabled CPUs can reach their lower power states for maximum power-savings and that also helps ensure other CPU cores have a larger thermal/power budget for hitting their rated turbo frequencies. That change/fix will be coming in the Linux 5.20 cycle later this summer.
With at least the initial Xeon Scalable "Sapphire Rapids" servers introducing Advanced Matrix Extensions (AMX), if the large register state of AMX isn't properly initialized it can result in CPU cores not hitting the lowest power states of the CPU cores. The Linux patches are about ensuring the AMX state is properly initialized before the Intel Idle driver so that the low-power idle states can be achieved: it's a difference of C1E versus the deeper C6 sleep state for cores.
This behavior of shallower sleep states if the AMX state isn't properly initialized is treated as an early implementation-specific behavior with Sapphire Rapids. It's certainly important though knowing this early behavior for AMX-enabled systems so the kernel properly deals with it for maximum power savings and ensuring the non-sleeping CPU cores have a greater power/thermal budget for hitting their rated turbo frequencies.
The news this week is the patches have landed in TIP's x86/fpu branch. Now past the v5.19 merge window, this Sapphire Rapids improvement will be coming later this summer in the Linux 5.20 kernel cycle.
With at least the initial Xeon Scalable "Sapphire Rapids" servers introducing Advanced Matrix Extensions (AMX), if the large register state of AMX isn't properly initialized it can result in CPU cores not hitting the lowest power states of the CPU cores. The Linux patches are about ensuring the AMX state is properly initialized before the Intel Idle driver so that the low-power idle states can be achieved: it's a difference of C1E versus the deeper C6 sleep state for cores.
This behavior of shallower sleep states if the AMX state isn't properly initialized is treated as an early implementation-specific behavior with Sapphire Rapids. It's certainly important though knowing this early behavior for AMX-enabled systems so the kernel properly deals with it for maximum power savings and ensuring the non-sleeping CPU cores have a greater power/thermal budget for hitting their rated turbo frequencies.
The news this week is the patches have landed in TIP's x86/fpu branch. Now past the v5.19 merge window, this Sapphire Rapids improvement will be coming later this summer in the Linux 5.20 kernel cycle.
Add A Comment