[Bug 107410] New: [SKL] Up to 20% performance regression in GpuTest Triangle, due to 2x higher CPU power usage

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=107410

            Bug ID: 107410
           Summary: [SKL] Up to 20% performance regression in GpuTest
                    Triangle, due to 2x higher CPU power usage
           Product: DRI
           Version: DRI git
          Hardware: Other
                OS: All
            Status: NEW
          Severity: normal
          Priority: medium
         Component: DRM/Intel
          Assignee: intel-gfx-bugs at lists.freedesktop.org
          Reporter: eero.t.tamminen at intel.com
        QA Contact: intel-gfx-bugs at lists.freedesktop.org
                CC: david.weinehall at intel.com,
                    intel-gfx-bugs at lists.freedesktop.org
     i915 platform: SKL

Between following drm-tip commits:
2018-06-16 13:01:59 UTC d4b21cf9ff "2018y-06m-16d-13h-00m-25s UTC integration
manifest"
2018-06-17 12:43:31 UTC e47233f783 "2018y-06m-17d-12h-42m-13s UTC integration
manifest"

There were following 3D benchmark performance regressions on SKL GT4e:
- 20% GpuTest Triangle (windowed & composited, half screen size)
- 10% GpuTest Triangle (FullHD fullscreen i.e. slower)
- 1-6% SynMark Batch[2-4]
- 1-4% GfxBench Tessellation (onscreen only), GpuTest Julia32 (windowed),
SynMark GSCloth & ShMapPcf (fullscreen)
  (There may also be smaller regression in SynMark CSCloth & ShMapVsm)

SKL GT2 shows regression in same cases, but they're somewhat smaller.

There were no improvements in any 3D benchmarks from this, but I noticed small
increase in SIMD CPU copy and large increase in SIMD CPU read performance. 
However, that was only on SKL GT2, not GT4e, so it may be unrelated.


When looking at the RAPL & CAGF data from before and after the regression on
SKL GT4e for the GpuTop fullscreen Triangle case:
1. *CPU power usage doubled*
2. GPU power use decreased a bit
3. GPU freq drops from (max) 950 to 800 Mhz, but temperature is OK
=> kernel has turned this use-case from GPU to CPU/TDP bound

SKL GT2 isn't TDP limited, and there one sees only effect 1), besides the FPS
drop, i.e. there the use-case became more CPU bound, but not enough to drop GPU
freq.


Interestingly, these regression are visible only on:
- SKL GT2  (i5-6600K)
- SKL GT4e (i7-6770HQ "SkullCanyon")

*not* on:
- KBL GT2  (i5-7500U)
- KBL GT3e (i7-7567U)
- SKL GT3e (i5-6260U)
- CFL-S (pre-production)

On those devices there were no changes in CPU power usage (nor performance).

-- 
You are receiving this mail because:
You are the assignee for the bug.
You are the QA Contact for the bug.
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-gfx-bugs/attachments/20180727/8ad951a0/attachment-0001.html>