RAPL quirks on Sapphire Rapids?

[connorimes]

Good afternoon,

I'm updating some RAPL-related code to support the new Sapphire Rapids CPUs (model 0x8F), which are documented in the latest Intel Software Developers Manual (December 2022).

The Linux kernel implements at least two PSYS domain quirks which aren't documented in the SDM.  The first specifies a fixed 1 J energy unit for the domain's energy counter [1] and the second specifies a different bit configuration for the power limit register [2].

The kernel is _usually_ a reliable reference, but it's not perfect, and I don't currently have a Sapphire Rapids system to test on. Can somebody confirm or refute these exceptions?

[1] https://github.com/torvalds/linux/commit/2d798d9f5967a3f134b1c55acbbe026c032e3429
[2] https://github.com/torvalds/linux/commit/931da6a0de5d620425af4425344259e6ff46b654

Thanks in advance,
-Connor

[Anthony_7]

At least two PSYS domain peculiarities that aren't described in the SDM are implemented by the Linux kernel. For the domain's energy counter, the first defines a fixed 1 J energy unit, and the second specifies a different bit configuration for the power limit register [2].

[connorimes]

Yes, those are the two commits I linked to.  So we have two conflicting sources of information.  There have been errors in the kernel's RAPL implementation before, so my question is whether the SDM or the kernel (or perhaps neither) are correct.

[selenaaeslai]

The Sapphire Rapids family includes 52 SKUs (see chart) grouped across 10 segments, inclusive of the Max series: 11 are optimized for 2-socket.