Is AMD potentially risky? (Threadripper)

[Diorn]

43 minutes ago, tricecold said:

hey guys, i was going to post a link here too, but I see that its already shared. To be honest Grain solver is a disaster on any CPU and definitely would not lunch it on CPU if a GPU is available. From experience I can say that Flip will scale as well as Pyro. 1950X has an average speed per core , so I don`t expect it to shine on smaller tasks also, however there is an FEM simulation in there with a some sop solver involved, although I believe pcOpen would perform a lot better than an attribute transfer. I can probably speed up this setup another %30 with lots of compile  sops where possible, but yet again it will be limited to sop operators.

However the purpose of the Benchmark is to simulate a production environment. The goal is how efficient a 1950X would be if you were to deliver to the client all the way to the renders as a freelancer, This CPU is a waste in a studio environment as an artist machine.

Diorn, I am however curious to see those numbers on all the tasks that are completed

I can also try a setup of some complex SOP operators one with and without compile sop to see the diffrence

 

Sure no problem, will post the results, i'll run it a couple of times to see if the numbers stay consistent.

One mention though: i will not let it go to the mantra rendering. Can't keep my workstation unavailable for a few days at a moment and we already have cinebench scores and some mantra tests. Speaking as a freelancer,and especially from this position, CPU rendering is not an option for me, i would not want the hassle of online renderfarms and i am trying to keep my setup compact, hence one workstation for all things, from modeling texturing to rendering to comp. I could not afford to leave freelance jobs to render for on CPU on a single computer this is why Redshift is highly preferable for me. This is my current production environment. I did work in a facility where i had to take care of a 30 PC vray render farm (running on backburner) and i am glad those days are over for the types of jobs i like to work on. So i will only post the numbers up untill the Mantra tasks for now.

As a personal opinion/conclusion to the other tests we had on this thread, as i previously stated it is more important to decide what are your personal needs, for an effects TD i would say TR is a win no question. Again speaking as a generalist freelancer the jobs i do are varried and i cannot guarantee i'll do Pyro all year or flip or even heay sim stuff.  Still it is comforting to see lower res sims go fast on the intel, so this means faster RnD with the expense of leaving it to compute more for the final rez sim once and when I get there.

I'm really interested for any more under the hood or behind the scenes info you have about this stuff and those SOPs setups you mention sound interesting, so far i've fund this thread very educating and it is good to find out more about how Houdini works. Also the benchmark scene you posted is really interesting to learn from especially the deadline management of the rops, so i am grateful for that!

Edited September 18, 2017 by Diorn

[tricecold]

I agree with you, there are lots of different types of tasks, I also didn`t expect everybody to ran the entire tasks, as it would take 2 3 days to complete the task.  I have also done some i7 vs slower but dual socket xeon pyro simulation tests for 200 frames, I7 was at frame 100 where dual xeon was at frame 50, but things turn around once voxel numbers increase  and dual xeon would finish the 200 frames almost 30 ahead of a single I7. In the end you would save time in a finished cache. 

I am a big fan of redshift also, but I don`t do render for freelance. IMHO ,in the end it comes to the price and possible future optimizations on softwares ability to multi-thread better.

[Guest tar]

1 hour ago, tricecold said:

 To be honest Grain solver is a disaster on any CPU and definitely would not lunch it on CPU if a GPU is available.

 

If OCL CPU worked on TR/7900x you should get ~50% increase on the grain solver.

[Diorn]

3 hours ago, tricecold said:

I agree with you, there are lots of different types of tasks, I also didn`t expect everybody to ran the entire tasks, as it would take 2 3 days to complete the task.  I have also done some i7 vs slower but dual socket xeon pyro simulation tests for 200 frames, I7 was at frame 100 where dual xeon was at frame 50, but things turn around once voxel numbers increase  and dual xeon would finish the 200 frames almost 30 ahead of a single I7. In the end you would save time in a finished cache. 

I am a big fan of redshift also, but I don`t do render for freelance. IMHO ,in the end it comes to the price and possible future optimizations on softwares ability to multi-thread better.

Yup we are moving closer and closer to proper multi-threadding support for sure! Just was not convinced if it is right around the corner or not so I personally made a bit of a trade-off for now

Edited September 18, 2017 by Diorn

[tricecold]

3 hours ago, marty said:

If OCL CPU worked on TR/7900x you should get ~50% increase on the grain solver.

OpenCL does work on these CPUs, and it would made gain, I just noticed that the grain solver is a perfect example for GPU, unfortunetly it is not the best solution for big farms, as not everybody has a GPU farm, next to their CPU farm.

[Guest tar]

18 minutes ago, tricecold said:

OpenCL does work on these CPUs, and it would made gain, I just noticed that the grain solver is a perfect example for GPU, unfortunetly it is not the best solution for big farms, as not everybody has a GPU farm, next to their CPU farm.

cool - how did you get it to work? the others tired and it failed. Can you post the

hgpuinfo -c

for your TR. thanks!

[tricecold]

Are you on linux or windows, on windows openCL drivers is installed out of the box, but on ubuntu you need to install them seperately 

[Guest tar]

they are on Windows but everything we tried didn't work.

[tricecold]

It's Abit confusing, the thing is by default openCL tick box will use your GPU not CPU, if you want OpenCL to be ran on CPU , there was a flag you need to put in your houdini.env which I don't remember

[Guest tar]

4 minutes ago, tricecold said:

It's Abit confusing, the thing is by default openCL tick box will use your GPU not CPU, if you want OpenCL to be ran on CPU , there was a flag you need to put in your houdini.env which I don't remember

yep and it doesn't work for TR or 7900x in our testing. You can test it by setting

 

HOUDINI_OCL_DEVICETYPE=CPU

 

in the environment file.

 

Thanks!

Edited September 19, 2017 by tar

[tricecold]

Will try in an hour and let you know

[Diorn]

11 hours ago, marty said:

If OCL CPU worked on TR/7900x you should get ~50% increase on the grain solver.

I got open CL for CPU to turn on at the office of a studio i work for, on the i7 6950x, was indeed as simple as setting the ENV file, this would not work at home on the 7900x. i'll try to run the grain benchmark from VFX ARABIA, all i have to do with oCL set for CPU, is turn on all the OCL ticks inside the DOP network, correct?

here are the numbers for the Grains scene from VFX Arabia, res 0.01:

  No oCL -12:16

CPU oCL -08:04  40%    faster than no OCL

GPU oCL -06:15  63.7% faster than no OCL  (gtx 1080)

Edited September 19, 2017 by Diorn
added scores

[tricecold]

10 hours ago, marty said:

yep and it doesn't work for TR or 7900x in our testing. You can test it by setting

 

HOUDINI_OCL_DEVICETYPE=CPU

 

in the environment file.

 

Thanks!

Hey, so it works here, 

I am on Ubuntu 16.04

cliinfo

ICD loader properties
  ICD loader Name                                 OpenCL ICD Loader
  ICD loader Vendor                               OCL Icd free software
  ICD loader Version                              2.2.8
  ICD loader Profile                              OpenCL 1.2
Device Name                                     GeForce GTX 970
  Device Vendor                                   NVIDIA Corporation
  Device Vendor ID                                0x10de
  Device Version                                  OpenCL 1.2 CUDA
  Driver Version                                  375.66
  Device OpenCL C Version                         OpenCL C 1.2 
 

[DaJuice]

@marty, sorry I'm on Win10 as well.

@Diorn, I will try to run your benchmark in the next couple of days.

[Guest tar]

@tricecold thanks for testing - that info indicates that the GPU is being used for OCL. Can you please try the following to show all OCL devices available:

 

hgpuinfo -l

for example on a MacMini it will show the Intel Chips being available for OCL CPU:

hgpuinfo -l
OpenCL Platform           Apple
Platform Vendor           Apple
Platform Version          OpenCL 1.2 (May 26 2017 12:59:48)
OpenCL Device             Intel(R) Core(TM) i7-2635QM CPU @ 2.00GHz
OpenCL Type               CPU
Device Version            OpenCL 1.2 
Frequency                 2000 MHz
Compute Units             8
Device Address Bits       64
Global Memory             16384 MB
Max Allocation            4096 MB
Global Cache              0 KB
Max Constant Args         8
Max Constant Size         64 KB
Local Mem Size            32 KB
2D Image Support          8192x8192
3D Image Support          2048x2048x2048

 

[tricecold]

no no, i tried it inside Houdini with the environment flag on and off and with opencl ticket on and off, its almost twice as fast with the flag on, without the flag on in houdini.env and openCL ticket , its more than twice the speed, i will run the command once i get back also

[Beatnutz]

On 9/18/2017 at 8:21 AM, Diorn said:

Yes, at 4.8 it was actually asking for more current than i was giving it for it to become stable under load. But more current would mean larger temps at 4.8ghz i was getting auto throttling from the CPU hardware fail-safe mechanisms. 8 cores at 4.7 and 2 at 4.6 runs indeed hot, but no throttling and no instability so far for renders or sims or any other of the benchmarks, and some of them run hotter(90+) some of them run cold (70+).

PS: if you don't mind, and have the time, could you please try out my sops/Bullet benchmark on your machines?

@Beatnutz still around for some more tests?

 

Yeah sure! Been trying to do some actual work for a few days

[Diorn]

11 minutes ago, Beatnutz said:

Yeah sure! Been trying to do some actual work for a few days

 

On 9/19/2017 at 8:06 PM, DaJuice said:

@marty, sorry I'm on Win10 as well.

@Diorn, I will try to run your benchmark in the next couple of days.

cool guys, thanks. Regards,

 

[DaJuice]

@Diorn, Okay PorcelainSops_V1 scene took 15m 28s. Same settings as before: TR 1950x, at 3.8GHz, memory at 3,200MHz.

[Diorn]

11 hours ago, DaJuice said:

@Diorn, Okay PorcelainSops_V1 scene took 15m 28s. Same settings as before: TR 1950x, at 3.8GHz, memory at 3,200MHz.

Thank you! updated the scores spreadsheet...