[osg-users] CPU Performance issues with AMD 2700 vs Intel Corei7 4770S

[Robert Osfield]

Hi All,

I have recently got a new dev machine, a AMD2700 + Geforce 2060
system. I wanted a quiet mulit-core system as well as to get some
experiment working with AMD chips as I've been using Intel Corei7
chips for the last 8 years.  The machine uses passive cooling for
everything except the graphics card so it's soo quiet you hardly know
it's on :-)

I've begun performance testing and was expecting most improvements
across the board for single threaded tasks, and big improvement on
highly threaded tasks like building.   I expected this based on a
number of online performance comparisons.  For example:

https://cpu.userbenchmark.com/Compare/Intel-Core-i7-4770S-vs-AMD-Ryzen-7-2700/m2123vs3957

    http://hwbench.com/cpus/amd-ryzen-7-2700-vs-intel-core-i7-4770s

When using the OSG I'm seeing lower framerates for almost all the test
models I regularly use, in particular the cull traversal can be 20-40%
slower.

Doing synthetic OSG + VSG scene graph creation, traversal tests shows
the OSG 32% slower on the AMD (running osggroups from the osg2vsg
project) while the equivalent VSG test  is ~14% faster on the AMD
(running vsgvisitor from the vsgExamples project).

Looking at the Linux perf stat utility it looks like the OSG runs for
AMD run are hitting up against 498 context-switches for the run of
osggroups, while the Intel just sees 24 context switches.  Branch
misses are 0.66% on AMD vs 0.37% on Intel.  Instructions per cycle are
all worse on the AMD too, 0.72 (AMD) vs 0.95 (Intel)
instrictions/cycle

The same equivalent tests with the VSG shows that the AMD has 56
context switches vs 2 on the Intel.  Branch misses are also still
higher on the AMD, but the instructions per cycle are now 2.13 (AMD)
vs 1.96 (Intel) instructions/cycle.

Another way of presenting this data, this OSG vs VSG tests runs a
testing of creation 11 deep quad tree, traversing every node in the
graph 10 times, then destroying the graph:

VSG Intel 0.615 secs  (VSG 6.33 x faster than OSG)
OSG Intel 3.895 secs

VSG AMD 0.555 secs (VSG 9.36 x faster than OSG)
OSG AMD 5.196 secs


Despite the raw creation, traversals and destruction being quicker on
the AMD with the VSG, the VSG framerates on scenes are worse on the
AMD vs Intel though, except where the 2060's better fill rate gives it
advantage.  Even when I take the VSG's traversals out of the equation
and just submitted the same command graph to render each frame the
frame rate is worse on the AMD vs Intel system, suggesting it's not
just the slower traversals that is at issue - the Vulkan driver and
hardware look to be struggling as well.

The two systems aren't identical software wise, the Intel system in on
Kubuntu 18.04 with gcc 7.3.0 vs Kubuntu 18.10 with gcc 8.2.0.  The
Intel system has NVidia 396.54.09 vs AMD system with NVidia 415.27.

I wouldn't have thought the OS, compile and drivers to make such stark
differences in performance.  The VSG's lighterweight design does seem
to help reduce the impact a bit and for some tests gives the AMD an
advantage, but it's still overall slower on the AMD.  The OSG sees a
huge hit on complex scenes on the AMD system.

So I'm perplexed what might be going on.  The perf stat report
suggests that the AMD is running at around 3.9-4GHz vs Intel running
at 3.8 to 3.GHz.  The perf stat results suggests that certain
workloads are really messing up the AMD's chip ability to efficiently
handle the load.

On one upside, the AMD's 8 cores/16 threads vs Intel's 4cores/8
threads certainly helps when compiling the OSG, but this is small
comfort as my job is far more focused on optimizing real-time
rendering performance than just getting code compiled.

I'd really like some feedback from the community as to what might be
going wrong on my AMD system.  Are others seeing significantly slower
CPU performance on AMD vs Intel for certain tasks?

Any suggestions as what tests to do next?  What to tweak?

Cheers,
Robert.

I have