High hardware latency on i9-13900E

[chrigum]

I'm trying to setup a Linux system to have "decent" realtime capabilities with PREEMPT_RT. I'm happy if the worst case latency is around 100us across all CPUs. I don't want to do CPU isolation if possible.

However, I have a very hard time with 13th Gen Intel(R) Core(TM) i9-13900E on a R680E chipset.

If I add some CPU load with "stress-ng --x86cpuid 32" and run "hwlatdetect --duration 5m", I get latencies of 200us and more. All CPUs can use their full frequency scaling range and are set with EPP=0 and EPB=0, hardware p-states are enabled (intel_pstate active).

I've attached the output of x86_energy_perf_policy and hwlatdetect. I checked on Debian 12 and Ubuntu 22.04 using the Debian 6.1 RT kernel.

I discovered by accident that cpuid triggers huge latencies on this CPUs. I checked 3 systems (all same board/cpu combination).

The same board works quite well with 12th Gen i9 and smaller CPUs like 12th Gen i3. Any ideas what is going on?

[Jocelyn_Intel]

Hello, @chrigum  

Thank you for posting on the Intel® communities.  

We understand that you have an inquiry about Intel® Core™ i9-13900E Processor. 

We have a forum for those specific products and questions so we are moving it to the Embedded Intel® Core™ Processors Forum to get answered more quickly. 

Best regards,  

Jocelyn M.   

Intel Customer Support Technician. 

[Diego_INTEL]

Hello @chrigum,

Thank you for contacting Intel Embedded Community.

I checked internally and found that the BIOS version or a configuration may be affecting, you may try updating to the latest BIOS and also try disabling PCIe ASPM Support. Also, you my try with Ubuntu or Debian communities to check if they have more information that may be of help.

Best regards,

@Diego_INTEL 

[chrigum]

Hello @Diego_INTEL 

Thank you for your answer. I've added the following two kernel boot options:

• pcie_aspm=off
• pcie_port_pm=off

However, the situation did not improve and I still see very high hardware latencies. Is this expected with Hyperthreading? Or is "weird behavior" to be expected when running "cpuid" ASM instruction?

I only experience this with the 13th Gen i9. The 12th Gen i9 does not show this behavior, nor does a i3-12100E. They all run on the same custom R680E board.

When combined with some GPU load (like working on the computer), I see latencies up to 1800us.

Concerning BIOS update: We use a custom board. What BIOS version should I be looking for? The microcode of the CPU is 0x119, if this is any help.

Regards,

Christoph

[Diego_INTEL]

Hello @chrigum,

Yes, the microcode version was useful to check, there is a newer version, you may try to update everything available from the BKC.

Document #781359 - Raptor Lake S Refresh Consumer and Corporate PR2 Best Known Configuration (BKC) Software Package (Microsoft* Windows 11 22H2– 64 bit) Platform Validation Guidance

Best regards,

@Diego_INTEL 

[chrigum]

Hello @Diego_INTEL 

Very recently, microcode version 0x11d has been released to https://github.com/intel/Intel-Linux-Processor-Microcode-Data-Files.

I updated my system to the new version (cat /proc/cpuinfo reports 0x11d), but hwlatdetect still reports very high hardware latencies with hyperthreading and cpuid.

Anything else I could try?

Thanks,

Christoph