AMD - Increased Render Time vs Intel

[Andrew_Wilkins1]

Hi everyone,

We've recently run some rendering testing on an Intel vs AMD cpu.

In one of the render tests which contained simulation from Houdini, we saw a large difference in render time favoring Intel (Compared to all other render tests)

Analyzing the render logs the only major difference in the logs was the "BVH_motion" between the two CPU.

Intel #1:

| log started Tue Nov  5 17:41:18 2019
| Arnold 5.2.2.1 [2dff6c07] windows icc-17.0.2 oiio-2.1.0 osl-1.11.0 vdb-4.0.0 clm-1.0.3.513 rlm-12.4.2 2019/01/10 05:42:47
| host application: MtoA 3.1.2.1 b1ad83a6 (master) Maya 2018
| running on STRN-High-01, pid=5420
|  1 x Intel(R) Core(TM) i9-7920X CPU @ 2.90GHz (12 cores, 24 logical) with 65214MB
|  Nvidia driver version 432.00
|  GPU 0: GeForce GT 710 @ 954MHz (compute 3.5) with 1024MB (930MB available)
|  Windows 8 Professional (version 6.2, build 9200)

| top session self-times by category 
|  BVH_motion::intersect                                              9:49.83 (37.06%) 
|   /obj/ASS_peak_mist/points                                         3:22.93 (12.75%) 
|   /obj/ASS_surface_mist/points                                      2:27.79 ( 9.29%) 
|   /obj/ASS_white_water_foam/points                                  2:04.25 ( 7.81%) 
|   /obj/ASS_ocean_Tank/polygons                                      1:54.68 ( 7.21%) 

Intel #2:

| log started Tue Nov  5 16:29:07 2019
| Arnold 5.2.2.1 [2dff6c07] windows icc-17.0.2 oiio-2.1.0 osl-1.11.0 vdb-4.0.0 clm-1.0.3.513 rlm-12.4.2 2019/01/10 05:42:47
| host application: MtoA 3.1.2.1 b1ad83a6 (master) Maya 2018
| running on Ultra1-1, pid=11300
|  2 x Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz (32 cores, 64 logical) with 65181MB
|  Windows 8 Professional (version 6.2, build 9200)

| top session self-times by category 
|  BVH_motion::intersect                                              6:09.04 (35.63%) 
|   /obj/ASS_peak_mist/points                                         2:03.37 (11.91%) 
|   /obj/ASS_surface_mist/points                                      1:33.26 ( 9.00%) 
|   /obj/ASS_white_water_foam/points                                  1:19.92 ( 7.72%) 
|   /obj/ASS_ocean_Tank/polygons                                      1:12.38 ( 6.99%) 

AMD:

| log started Mon Nov 11 14:18:22 2019
| Arnold 5.2.2.1 [2dff6c07] windows icc-17.0.2 oiio-2.1.0 osl-1.11.0 vdb-4.0.0 clm-1.0.3.513 rlm-12.4.2 2019/01/10 05:42:47
| host application: MtoA 3.1.2.1 b1ad83a6 (master) Maya 2018
| running on TFRN-Ryzen, pid=6632
|  1 x AMD Ryzen Threadripper 2990WX 32-Core Processor (32 cores, 64 logical) with 65426MB
|  Windows 8 Professional (version 6.2, build 9200)

| top session self-times by category 
|  BVH_motion::intersect                                             15:03.11 (41.93%) 
|   /obj/ASS_peak_mist/points                                         5:11.03 (14.44%) 
|   /obj/ASS_surface_mist/points                                      3:45.58 (10.47%) 
|   /obj/ASS_white_water_foam/points                                  3:17.32 ( 9.16%) 
|   /obj/ASS_ocean_Tank/polygons                                      2:48.94 ( 7.84%) 

Question 1: What does "BVH_motion::intersect" mean?

Question 2: Why would there be such a big difference between the CPU architecture in this area?

Any info would be great!

Thanks very much,

Andrew

MtoA 3.3.0
Arnold Core 5.4.0.0
Maya 2018.2

[Stephen.Blair]

BVH is bounding volume hierarchy, an acceleration structure. Based on that name, here Arnold is tracing intersections through the BVH for motion blur.

Can you post the log headers too, so we can see what the two machines are like?

[Andrew_Wilkins1]

Thanks Stephen, I updated the main post with the machine info. (To keep all the data in one place 😉

Motion blur is disabled on the scene, but let me triple check that. Would if matter if motion blur was enabled when the ASS files were exported?

[Andrew_Wilkins1]

Hi Stephen,

There is actually a chance that motion blur was enabled at the time of render testing. We're planning to re-run the test.

I will update you on that.

Do you think that motion blur calculation could being playing a part here?

[Stephen.Blair]

I'm not sure, we need to get one of the core developers to answer that.

Arnold doesn't do anything special for AMD or Intel, so it's not any optimization on our part for Intel

[Stephen.Blair]

ThreadRipper has worse memory bandwidth than Intel. To rule that out, repeat the test with 1 thread to see if it's still twice as slow. If it was memory bandwidth, AMD will now run at about the same speed as Intel since there's now 32x less memory being read.

If that's the issue, then first thing to check is that you have the correct memory DIMMs in the AMD machine. We've had issues in the past with customers not having enough DIMMs or not the right speed DIMMs.

How many DIMMs do you have in the AMD machine, and what model they are?

Here's what you should have to get quad channel:
https://en.wikichip.org/wiki/amd/ryzen_threadripper/2990wx#Memory_controller

Note how you need exactly 4 or 8 dimms and you need to place them in the right slots. Also, note the DDR4 speeds

[Andrew_Wilkins1]

Thanks so much for the info,

I'm following up atm and will post findings here.

[Andrew_Wilkins1]

Hi again,

We haven't run the "1 thread" test yet, but I did follow up on the RAM and configuration.

It's currently running 4x 16GB 2666MHz, see attached image for config.
The RAM seems to running at 2400MHz, but we're looking into that—might just be a BIOS setting.

Regards,

Andrew

[Andrew_Wilkins1]

Hi,

Some more info here—we tested with and without motion blur and concluded that motion blur isn't the difference in render time.

Here are the additional "Thread Count" tests results—with only the "mist points" ASS files. As these are the assets in the scene which are rendering slower on the Ryzen compared to Intel. (Points exported from Houdini as an ASS file)

Single Thread

1 x AMD Ryzen Threadripper 2990WX 32-Core Processor (32 cores, 64 logical) with 65426MB

| frame time                2:15:07.71    machine utilization (99.90%) 
| top session self-times by category 
|  BVH_motion::intersect                                       51:01.24 (37.71%) 
|   /obj/ASS_surface_mist/points                                     27:57.37 (20.66%) 
|   /obj/ASS_peak_mist/points                                        23:03.87 (17.05%) 
|  points::intersection                                              21:33.25 (15.93%) 
|   /obj/ASS_surface_mist/points                                     13:20.13 ( 9.86%) 
|   /obj/ASS_peak_mist/points                                         8:13.11 ( 6.07%) 

All Thread

1 x AMD Ryzen Threadripper 2990WX 32-Core Processor (32 cores, 64 logical) with 65426MB

| frame time                   9:34.67    machine utilization (98.98%) 
| top session self-times by category 
|  BVH_motion::intersect                                              4:13.26 (44.13%) 
|   /obj/ASS_surface_mist/points                                      2:17.35 (23.93%) 
|   /obj/ASS_peak_mist/points                                         1:55.91 (20.20%) 
|  points::intersection                                               1:49.65 (19.11%) 
|   /obj/ASS_surface_mist/points                                      1:06.79 (11.64%) 
|   /obj/ASS_peak_mist/points                                         0:42.86 ( 7.47%) 

---

Single Thread

2 x Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz (32 cores, 64 logical) with 65181MB

| frame time                3:14:25.51    machine utilization (99.96%) 
| top session self-times by category 
|  BVH_motion::intersect                                           1:15:49.48 (38.94%) 
|   /obj/ASS_surface_mist/points                                     39:27.65 (20.27%) 
|   /obj/ASS_peak_mist/points                                        36:21.83 (18.68%) 
|  points::intersection                                              30:06.10 (15.46%) 
|   /obj/ASS_surface_mist/points                                     19:22.22 ( 9.95%) 
|   /obj/ASS_peak_mist/points                                        10:43.87 ( 5.51%) 

All Thread

2 x Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz (32 cores, 64 logical) with 65181MB

| frame time                   5:39.37    machine utilization (97.09%) 
| top session self-times by category 
|  BVH_motion::intersect                                              2:04.37 (36.72%) 
|   /obj/ASS_surface_mist/points                                      1:06.72 (19.70%) 
|   /obj/ASS_peak_mist/points                                         0:57.65 (17.02%) 
|  points::intersection                                               0:54.24 (16.01%) 
|   /obj/ASS_surface_mist/points                                      0:33.99 (10.04%) 
|   /obj/ASS_peak_mist/points                                         0:20.24 ( 5.98%) 

Thanks!

Andrew