Intel Xeon Phi 7210 Frequency

[CPati2]

Hi All,

I am using Intel's Xeon Phi 7210 model of x200. It seems to be that the core frequencies are hard set to 1.3GHz. Can anyone please confirm if they have similar CPU?

This is after turbostate showing consistent ~1.3GHz as the frequency for all 64 cores. I also logged data after setting scaling_max_freq and scaling_mix_freq to 1.5GHz, but the cur_freq sysfs keeps showing ~1.3GHz only. I am running performance governor and other option is powersave due to inter_pstate driver.

Thanks.

[jimdempseyatthecove]

Mine reports 1.3GHz too. I haven't (and won't) play with overclocking.

While in your case you may gain 15% improvement (0.2 / 1.3), one typically gets much better performance by writing better code.

Jim Dempsey

[McCalpinJohn · Accepted solution]

Processor frequency information for the Xeon Phi x200 series is contained in a rather unusual place -- the "product brief" document:  https://www.intel.com/content/www/us/en/processors/xeon/xeon-phi-processor-product-brief.html

From footnote 3 on page 4, it looks like the rules are:

• The nominal frequency is for all cores in use.
• The maximum Turbo frequency for all tiles active is 0.1 GHz above the nominal frequency.
• The maximum Turbo frequency for one tile active is 0.2 GHz above the nominal frequency.
• The frequency drops by 0.2 GHz if there is a high proportion of AVX instructions.

My experience on the Xeon Phi 7250 (1.4 GHz nominal) is mostly consistent.  There are some cases were we don't see the 0.2 GHz reduction for code that contains a high proportion of 512-bit AVX-512 instructions, but we don't consider that to be a problem... ;-)

The range of available frequencies is so small that I have not looked at whether there are differences in maximum Turbo frequency for 128-bit, 256-bit, and 512-bit vector use.

[JJK]

I've got a Xeon Phi 7210 (Ninja dev platform) running at 1.3 GHz nominal; the default BIOS setting on this supermicro box is that turbo boost is DISabled; you can enable it in the BIOS under "Advanced Power Management Configuration". First enable "EIST (P-states)" then select the option "CPU P state control" and enable Turbo Boost.

Having said that, TurboBoost does not yield fantastic results on the Phi, even for single core jobs - I was hard pressed to get any job running at something close to 1500 MHz.  If you disable HyperThreading you can reach ~ 1480 MHz for  a single core job, but what's the point in doing that? 

[CPati2]

Hi All,

I am setting scaling_min_freq and scaling_max_freq, but it looks like the system isn't accepting the values I am setting. Turbo mode is ON via BIOS.

$> sudo cpupower frequency-set -u 1000000

$> sudo cpupower frequency-set -d 1000000

$> sudo cpupower frequency-info
analyzing CPU 0:
  driver: intel_pstate
  CPUs which run at the same hardware frequency: 0
  CPUs which need to have their frequency coordinated by software: 0
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 1000 MHz - 1.50 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 1000 MHz and 1000 MHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency: 1.31 GHz (asserted by call to hardware)
  boost state support:
    Supported: yes
    Active: yes

I think current CPU frequency should be 1 GHz?

Thanks.

[McCalpinJohn]

Two issues:

1. The "current CPU frequency" reported by "cpupower frequency-info" is not a reliable way to determine the frequency that will be used for long-running jobs.   I recommend running "perf stat" on a real program with a running time of more than a few seconds to see what the governor actually does.  In my experience on other Intel processors, long-running jobs run at the requested frequency whenever that is possible. 
2. The power control unit in (PCU) Intel processors sometimes chooses frequencies higher than what you have requested.  This can be influenced by other internal parameters set by the use of the "performance" governor.  For most Intel Xeon processors, the uncore performance monitoring manual provides information on PCU events.  Unfortunately the KNL performance monitoring documentation does not describe any PCU events.

Moving forward, Skylake Xeon systems include a whole new infrastructure for frequency control that (when enabled) overrides all of the legacy controls.  This is called "Hardware-Controlled Performance States" (HWP), and is enabled by default on our Xeon Platinum 8160 processors.  The "cpupower" utility interacts with HWP, but there are bugs in the CentOS/RHEL 7.3 implementation of this interaction, so I had to write a tool to write the HWP control registers directly to get the level of control that used to be available using a the "userspace" cpufreq governor for earlier systems....