[arch3990]

Hello,

I use Maxwell Render on AMD and Intel systems. I examined a performance problem with certain Maxwell scenes on the AMD system and I can't figure out, what the problem is...

Along you will find different Maxwell scenes with Benchmark results of three different rendering Systems for better comparison.

System Specs:
- Intel Core i7-2600K @ 5.0 GHz / 4GB / Win7 X64
- 2x Intel Xeon X5680 @ 4.20 GHz / 12GB / Win7 X64
- 4x AMD Opteron 8439SE @2.80GHz /16GB / Windows 2003 X64



"Benchwell" - Resolution: 960x768 - SL 15

2x X5680: 2m29s / 4280.58
4x 8439SE: 3m09s / 3381.77 - ok
i7-2600K: 5m36s / 1903.18



"Clabo 356" - Resolution: 1600x1200 - SL 16

2x X5680: 24m31s / 1710.31
4x 8439SE: 41m13s / 1017.75 - seems ok - should be a little bit faster
i7-2600K: 54m46s / 764.44



"HOVAL" - Resolution: 1600x1200 - SL 16

2x X5680: 43m49s / 957.40
4x 8439SE: 52m50s / 794.08 - ok
i7-2600K: 1h43m29s / 259.32



"BOBECK" - Resolution: 1280x960 - SL 16

2x X5680: 55m21s / 485.12
i7-2600K: 1h44m31s / 256.80
4x 8439SE: 2h05m25s / 214.18 - much too slow


"Mini-Football" - Resolution: 1024x768 - SL 15

2x X5680: 4m58s / 2294.41
i7-2600K: 8m24s / 1354.04
4x 8439SE: 14m31s / 787.59 - much too slow

...other Benchmark - Results of the systems:

Cinebench 11.5:
2x X5680: 21.21
4x 8439SE: 17.82
i7-2600K: 9.78

wprime 1024m:
2x X5680: 1m02.033s
4x 8439SE: 1m17.642s
i7-2600K: 2m34.157s

The AMD (8439SE) system must be nearly 2x faster than the single 2600K system but in several Maxwell Scenes it is much slower...

Does anybody have an idea what might be the reason for the big differences in Render Times in relationship to the computing power in some of the scenes?

[dmeyer]

Do the scenes where it appears slow have displacement?

My first guess is that the lower per-core speed of the AMD system is what is impacting it in some scenes where there is a rendering step that is single threaded, such as displacement. You can watch this by keeping an eye on Activity Monitor/Task Manager during the render process.

[Mihai]

In Maxwell, File>Preferences>Render, set Min. time for saving to disk to something like 15min and check 'save to disk only at end'. This way you rule out any differences in disk writing speeds. It may be that for the very simple renders, it's pausing the render while writing quite often to disk the MXI and the AMD machine has slower disks?

[arch3990]

Do the scenes where it appears slow have displacement?

...no displacement in the scenes

In Maxwell, File>Preferences>Render, set Min. time for saving to disk to something like 15min and check 'save to disk only at end'. This way you rule out any differences in disk writing speeds. It may be that for the very simple renders, it's pausing the render while writing quite often to disk the MXI and the AMD machine has slower disks?

I tried "save to disk only at end" and the result with the "Mini-Football" - Scene and the Opteron-System is: 14m29s / 789.40 - 2s faster
Normaly I save on a mapped Network drive (Gigabit-Lan / 6x Raid-0) which is quite fast - so all Render Systems "share" the same speed / performance. For the tests above I used the local HDD, which is a simple Samsung HD252HJ (using AHCI) on each system.


After a little research I found out something interesting:
I started examining the materials in the "Football" scene and I recognized, that there was the "R2" option set active in both materials "background" and "football" - don't ask me why

I did some further testing:


"R2" enabled in materials "background" + "football"
2x X5680: 4m58s / 2294.41
i7-2600K: 8m24s / 1354.04
4x 8439SE: 14m31s / 787.59 -

"R2" disabled in material "background" and enabled in material "football"
2x X5680: 3m48s / 2994.23 - 1m10s better
i7-2600K: 8m17s / 1375.73 - 7s better
4x 8439SE: 8m28s / 1348.73 - 6m3s better -

"R2" disabled in materials "background" + "football"
2x X5680: 3m31s / 3233.77 - 1m27s better
i7-2600K: 8m06s / 1403.83 - 28s better
4x 8439SE: 6m18s / 1810.74 - 8m13s better -



I think the AMD system doesn't like the "R2" option in the Material Editor and it is a "big handbrake" (in comparison with the Intel Systems)...

Now I have to find out what the problem might be with the other scene ("Bobeck")

[dmeyer]

Well thats interesting...

[zdeno]

GREAT FIND arch3990

ten gallons of beer is minimum NL should reward You

very very interesting.
I am curious how long this "bug/feature" is with us

[arch3990]

GREAT FIND arch3990

ten gallons of beer is minimum NL should reward You

very very interesting.
I am curious how long this "bug/feature" is with us

Thank you - and the story continues:

Where is the performance of my AMD System in Maxwell - Part III or "Intel vs. AMD in the Maxwell Arena - the battle goes on..."

I examined the "Bobeck" Scene because the AMD - System (24x 2.80 GHz) looses nearly 30 minutes on the Intel 2600K System ("8"x 5.0 GHz). For the background material in the original scene I used the wizard and made a plastic material which looked like that:


"Shininess" - 90% / "Reflectance" - 5% / Color: "Reflectance 0" - 230, 230, 230 / "Reflectance 90" - 220, 220, 220

Results:
1280x960 - SL 16
2x X5680: 55m21s / 485.12
i7-2600K: 1h44m31s / 256.80
4x 8439SE: 2h05m25s / 214.18 - 28m50s slower than 2600K



...same scene but background material like that:


"Shininess" - 30% / "Reflectance" - 30% / Color: "Reflectance 0" - 230, 230, 230 / "Reflectance 90" - 220, 220, 220

Results:
1280x960 - SL 16
2x X5680: 51m21s / 523.40 - 4min faster
4x 8439SE: 1h08m51s / 390.03 - 56min and 34sec faster!!!!
i7-2600K: 1h59m20s / 224.88 - 14m49s slower

I ran the tests three times on each machine and took the best results.

Conclusion:

- Maxwell doesn't like "shiny" materials on AMD - Systems
- Maxwell does like "shiny" materials on Intel - Systems

...or why is the 2600K so much faster with the shiny background material?

Question:
Is there any way to get rid of that limitation?

[Mihai]

This seems very strange to me, I doubt very much it is due to this material. Did you run the test with the same material several times to get an average? Is there something else on the machine that could be slowing it down from time to time? AV, scheduled tasks etc? How often do these machines write to disk for these tests?

[arch3990]

This seems very strange to me, I doubt very much it is due to this material. Did you run the test with the same material several times to get an average? Is there something else on the machine that could be slowing it down from time to time? AV, scheduled tasks etc? How often do these machines write to disk for these tests?

I ran the tests three times on each machine and took the best results.

The scene was the same on every machine - I ran the tests 3 times on every machine for both background materials (6 times).
Except for the big difference between the times for the two materials the times on each of the three tests using the same materials differed only a few seconds and in the end I took the best results of every machine. On the AMD System I have to use Windows 2003 X64 because of the limitation of CPU sockets in Windows 7 X64 but I recognized that "2003" is even a little bit faster than "7" (I think because there aren't running so much tasks in the background...) On the render Systems there aren't running any needless tasks or programs in the background and except for the differences in the specific hardware architecture (CPU-Mainboard-RAM) I use standard hardware on all machines (VGA: FX370 / HDD: HD252HJ).

For a better comparison I even did a test with the X5680 "downclocked" to 2.80 GHz (same Frequency than the AMD-System). It is optimized in comparison to a normal Xeon HEX-Core Setup because it runs with 14x 200 Bus Speed and 1000 MHz DRAM Frequency. Both Systems are using 24 Threads.

"Benchwell"
2x X5680 @2.80 GHz: 3m39s / 2921.67
4x 8439SE @2.80 GHz: 3m09s / 3381.77 - faster!



"Mini-Football" without "R2" in the materials and with "white" (0°=230/90°=220) background - "Shininess"=30 and "Reflectance"
2x X5680 @2.80 GHz: 6m49s / 1674.00
4x 8439SE @2.80 GHz: 6m15s / 1825.17 - faster!



"Clabo 356" - original scene without modification
2x X5680: 35m51s / 1170.00
4x 8439SE: 41m13s / 1017.75 - slower! (because of a lot of shiny "glas - material" ???)



"Bobeck" without "R2" in the materials and with "white" (0°=230/90°=220) background - "Shininess"=30 and "Reflectance=30"
2x X5680 @2.80 GHz: 1h15m18s / 356.63
4x 8439SE @2.80 GHz: 1h08m51 / 390.03 - faster!



"Cinebench 11.5"
2x X5680 @2.80 GHz: 14.21
4x 8439SE @2.80 GHz: 17.82 - faster!



"wprime 1024m"
2x X5680 @2.80 GHz: 1m32.859s
4x 8439SE @2.80 GHz: 1m17.642s - faster!



So you can see, that the per Core Performance of the AMD System is even better than the per Core Performance of the Intel Xeon DP Hex-Core System.

I noticed this problem for a long time now but I never had the time to find out why. When I use "shiny" materials like glas or shiny plastic in my Maxwell - Scenes, Rendering was always very slow on the AMD System (I could give you a lot more examples...) As you can see, if you "optimize" the scene (not using "R2" or "shiny" material) the AMD system is even faster than the Intel system (if using the same clock speed).

[juan]

I think the AMD system doesn't like the "R2" option in the Material Editor and it is a "big handbrake" (in comparison with the Intel Systems)...

Very interesting tests, thanks. When R2 is enabled there are not additional memory allocations but just a few more calls to the pow(float, float) function. It is surprising performance changes so dramatically in that test. It would be needed to make more tests to confirm that AMD pow() is remarkable slower, than Intel's but so far it looks so. Besides to this, tests can be done using just one thread, to check if it has something to do with sharing issues in the processor unit.

[dmeyer]

I've got AMD and Intel systems of almost every vintage from the past 4 years or so. If you'd like, post some scenes and I can verify or see if it may be an issue particular to the 84xx series. I don't believe we have any 8439's exactly but I've got other Istanbul core machines that should behave similarly.

[Aji Enrico]

Perhaps this is farfetched, but anyone remember the old "intel compiler crippled amd (and VIA) processors"? Maybe they have fixed this by now, didn't really follow up on that story myself.

Here's a little google search on it.
http://www.osnews.com/story/22683/Intel ... m_Compiler_
http://www.agner.org/optimize/blog/read.php?i=49 (scroll down a bit)

[arch3990]

Very interesting tests, thanks. When R2 is enabled there are not additional memory allocations but just a few more calls to the pow(float, float) function. It is surprising performance changes so dramatically in that test. It would be needed to make more tests to confirm that AMD pow() is remarkable slower, than Intel's but so far it looks so. Besides to this, tests can be done using just one thread, to check if it has something to do with sharing issues in the processor unit.

I think that it shouldn't be, that a system which is in most rendering applications as fast as a X5680 DP System (@ stock) slows down on a level of a Core i7-920 @stock (2.66 GHz). I can compare it because this was the CPU I used to have in one of my editing-systems and sometimes MX - Renders on the 8439 took as long as with the 920 (or longer)...

I've got AMD and Intel systems of almost every vintage from the past 4 years or so. If you'd like, post some scenes and I can verify or see if it may be an issue particular to the 84xx series. I don't believe we have any 8439's exactly but I've got other Istanbul core machines that should behave similarly.

Thank you - that would be nice - here we go:

"Football" Scene - original:
https://rapidshare.com/files/1603472925 ... iginal.zip
"Football" Scene - AMD optimized (no "R2"):
https://rapidshare.com/files/1986047373 ... imized.zip


"Bobeck" Scene - original:
https://rapidshare.com/files/524531321/ ... iginal.zip
"Bobeck" Scene - AMD optimized (Background - S-30% / R-30% instead of S-90% / R-5%):
https://rapidshare.com/files/3291526224 ... imized.zip

[dmeyer]

"Football" Scene - original:
https://rapidshare.com/files/1603472925 ... iginal.zip
"Football" Scene - AMD optimized (no "R2"):
https://rapidshare.com/files/1986047373 ... imized.zip


"Bobeck" Scene - original:
https://rapidshare.com/files/524531321/ ... iginal.zip
"Bobeck" Scene - AMD optimized (Background - S-30% / R-30% instead of S-90% / R-5%):
https://rapidshare.com/files/3291526224 ... imized.zip

Download parameters invalid - please try again later.

Will PM you with an upload location...

[dmeyer]

I have started these scenes on the following systems:

System A: 2x Opteron 2427 | 2.2 Ghz | 12 total cores | 8GB RAM | Win 7 Pro
System B: 2x Xeon X5650 | 2.66 Ghz | 12 physical, 24 logical cores | 12GB RAM | Win 7 Pro
System C: 4x Opteron 8354 | 2.2 Ghz | 16 total cores | 16GB RAM | Win Server 2008

Scene: "Bobeck Original"

System A: 2:21:25 @ 189.858
System B: 1:16:44 @ 350.048
System C: 2:11:45 @ 203.883


Scene: "Bobeck AMD Optimized"

System A: 2:23:52 @ 186.713
System B: 1:13:59 @ 363.058
System C: 1:53:48 @ 236.040


Scene: "Volley Mini Football Original"

System A: 13:33 @ 844.715
System B: 7:12 @ 1588.846
System C: 14:26 @ 794.048


Scene: "Volley Mini Football AMD Optimized"

System A: 9:42 @ 1179.701
System B: 4:53 @ 2341.314
System C: 8:17 @ 1381.258

Only scene changes I made are I turned off multilight and MXI and PNG outputs that were specified to eliminate disk system variables.

All tests done on Maxwell 2.6