[Intel-gfx] Iris pro broadwell power issue

[Joe Bloggsian]

Hello,

Our application uses intel HD graphics and media hardware. To date we
have been successfully using ivy bridge and haswell processors.
Recently we have been evaluating iris pro broadwell in the form of
E3-1265Lv4 and to our surprise have found that it seems to use much
more energy (~5-10x) for the same amount of work done compared to
other devices, e.g. ivy bridge i7 or non-iris pro broadwell. We see
the same in various test cases e.g. a simple EU shader that copies
rectangular video, the H264 HW decode, or surface scaling (through
vaapi). Here is a graph showing power usage for H264 video decode,
i.e. we are comparing the power usage of the two devices vs workload:

http://picpaste.com/pics/decoder_power-xpY94kKI.1458660426.jpg

Increased power is seen for the PP1/GPU domain only (read with MSRs,
confirmed with turbostat). No issue with CPU workloads. We are using
the same code with Ubuntu 15.10 in all cases. Our current assumption
is that there must be some configuration issue with the iris pro
broadwell xeon as it's very hard to conceive that adding the L4 cache
should suddenly 10x the power...

Wondering if anyone has any ideas of what could be causing this (badly
set MSRs/GPU regs, or L4 cache config, maybe?)

Cheers,
Joe