Jump to content
kemijo

Is AMD potentially risky? (Threadripper)

Recommended Posts

So what's the implication of disabling Core Performance Boost in Houdini? Is it still using the full power of ThreadRipper?

Share this post


Link to post
Share on other sites

I believe Core Performance Boost can dynamically apply a minor overclock across all cores and boost four cores up to 4.2GHz (In AMD lingo: "Precision Boost" and "XFR" respectively). Somebody more knowledgeable please correct me if I'm wrong.

Couple of Cinebench numbers:

CPB on: 2961

CPB off: 2938

Pretty minor dip but that's probably not representative of more single-threaded applications, where that 4.2GHz boost would be more helpful. If you overclock the processor it looks like CPB is disabled regardless. At the moment I have it overclocked to 3.75GHz across all cores (3199 in Cinebench).

Edited by DaJuice
  • Like 1

Share this post


Link to post
Share on other sites

@DaJuice and others, any tests that I can run to compare the TR to Ryzen 1700? I'm expecting TR to be twice as fast.  Thx!

Share this post


Link to post
Share on other sites

EDIT - Link to the boolean sphere file, in case anyone has trouble downloading the attachment: https://drive.google.com/open?id=0B0BbxninBS49aEVYa21TR3JYclU

-----------------------------------------------------

Hi Marty, I don't know if you have anything specific in mind but here's a quick scene I threw together that renders a short 45-frame sequence. A dense poly sphere that gets booleaned, converted to VDB, back to polys and rendered via Mantra. Open a Render Scheduler tab if you don't have one open and hit "Render to Disk in Background" in the Mantra ROP. The work is split roughly 30/70 between SOP cooking and rendering.

This machine completed the sequence in 15min 53sec (1950x @ 3.75GHz, Memory @ 3200MHz). Curious what you get on your machine, and also feel free to share any scenes you want me to run.

booleanSphere.hip

booleanSphere.jpg

Edited by DaJuice
  • Like 3

Share this post


Link to post
Share on other sites
On 2017/08/29 at 8:24 PM, DaJuice said:

This machine completed the sequence in 15min 53sec (1950x @ 3.75GHz, Memory @ 3200MHz).

Bang-on the prediction - 31min 30 sec. Ryzen 1700. TR is super nice!

UPDATE - just set the bios to 'performance',  the GHz goes from 3 to 3.6 or something, now it renders in 20min 20sec...  both times though the render stopped part way through. Not so good as the Render Manager showed 100% completed... hmmm.

Edited by marty

Share this post


Link to post
Share on other sites

Oh boy am I glad I found this thread. Thank you!

I recieved my Threadripper 1950X and MSI X399 motherboard yesterday and everything works fine except Houdini. I've been going nuts about it.
I'll try disabling the Core Performance Boost too. But what also helped me a bit was to use the AMD Ryzen Balanced power plan. Sometimes the simulations will work with no crash, but I still see crashes sometimes.

Of the few tests I've done so far performance with the TR is really bad compared to my old 5820K. Maybe it will be solved now. I'll get back to u on this.

Edited by Beatnutz

Share this post


Link to post
Share on other sites

Yeah it worked!! I'm so relieved!

Still not happy about simulation times though. Most of my benchmarks score lower (by a lot) than my 5820K so thats quite disappointing considering I bought the TR just for Houdini. Pyro is the one that seems fastest.

Edited by Beatnutz

Share this post


Link to post
Share on other sites

Hi guys, been weighing between TR and I9 7900x all summer.  Waited for the TR to arrive to see some benchmark scores and have orderd a 7900x this week.

AMD has done great, but one thing to take into account in Houdini is single thread operations wich are not few, 7900x will always beat TR in this department, but an overclocked 7900x can get close to TR multithread performance, at least in benchmark scores. I should be getting all the parts and have the rig running by next week. Eager to try DaJuice's benchmark. Will post my findings here for future reference for anyone else trying to make the decision.
 
Best,
Dorin

Share this post


Link to post
Share on other sites
3 hours ago, Diorn said:

Hi guys, been weighing between TR and I9 7900x all summer.  Waited for the TR to arrive to see some benchmark scores and have orderd a 7900x this week.

AMD has done great, but one thing to take into account in Houdini is single thread operations wich are not few, 7900x will always beat TR in this department, but an overclocked 7900x can get close to TR multithread performance, at least in benchmark scores. I should be getting all the parts and have the rig running by next week. Eager to try DaJuice's benchmark. Will post my findings here for future reference for anyone else trying to make the decision.
 
Best,
Dorin

A very good choice, I don't think you'll be disappointed.

I've had some time to test TR in Houdini today and pyro is the only thing it is good at. It falls short at pretty much everything else. I'm sure 7900x will crush TR in Houdini by a margin in single & multi threaded operations.

Edited by Beatnutz

Share this post


Link to post
Share on other sites
1 hour ago, Beatnutz said:

I'm sure 7900x will crush TR in Houdini by a margin in single & multi threaded operations.

Benchmarks please, not 'feelings'.  Please post up the tests and .hip files for others to check. Thanks!

Tbh surprised you're hitting a lot of single threaded bits. Houdini is very well threaded so it would be good to see which nodes are slowing you down.

 

EDIT: tests show that the 7900x will be faster. so follows with Houdini:

http://wccftech.com/amd-ryzen-threadripper-1950x-cpu-performance-benchmarks-leak/

Edited by marty

Share this post


Link to post
Share on other sites
17 hours ago, marty said:

Benchmarks please, not 'feelings'.  Please post up the tests and .hip files for others to check. Thanks!

Tbh surprised you're hitting a lot of single threaded bits. Houdini is very well threaded so it would be good to see which nodes are slowing you down.

 

EDIT: tests show that the 7900x will be faster. so follows with Houdini:

http://wccftech.com/amd-ryzen-threadripper-1950x-cpu-performance-benchmarks-leak/

Super simple scenes really. Wish I had planned it out better but I figured TR will blow my socks off and then I won't care too much about the benchmarks anyway.
I've received complaints that I should have cached the sim first and use OpenCL CPU etc. Too late now. But if you have a HIP you want me to test, send one over.

All scenes exported to Alembic (fram 1-240).
 


 

Results.jpg

Fluid.hip

Grain.hip

Pyro.hip

Share this post


Link to post
Share on other sites

Don't like these uber basic simple test but Fwiw

Grain: 1min 43sec

Pyro: 9min 12sec

Flip: 1min 30sec


Ryzen 1700 @ 3.6GHz - OpenCL CPU on. fileCaches written to M2. . Ubuntu 17.04 - 64Gb ram.

Edited by marty

Share this post


Link to post
Share on other sites

I'll do another test with caching and OpenCL tomorrow when I wake up. Feel free to post a different test if you want me to run it 

Share this post


Link to post
Share on other sites

try gamemode too - it turns off half the cores - should be in the bios

Share this post


Link to post
Share on other sites

I don't think IPC has changed much in the last few generations, so the 5820k should still perform great in single-threaded ops compared to processors with tons of cores that are clocked lower.

@Beatnutz, are you able to upload the hip files somewhere else? I'd like to give it a whirl but the forum is telling me I don't have access to them for some reason.

Share this post


Link to post
Share on other sites

Okey, so now counting the cache time and with OpenCL set. Not a very big gain with OC, how do I know if I'm using CPU or GPU?

Grain
1950x (old) - 2:21min
1950x - 1:51min
1950x @ 3.8GHz - 1:48min

Pyro
1950x (old) - 7:37min
1950x - 6:35min
1950x @ 3.8GHz - 6:32min

Fluid
1950x (old) - 2:07min
1950x - 0:32min
1950x @ 3.8GHz - 0:30min

Share this post


Link to post
Share on other sites
16 hours ago, DaJuice said:

I don't think IPC has changed much in the last few generations, so the 5820k should still perform great in single-threaded ops compared to processors with tons of cores that are clocked lower.

@Beatnutz, are you able to upload the hip files somewhere else? I'd like to give it a whirl but the forum is telling me I don't have access to them for some reason.

I couldn't download your file either. .. strange.
I uploaded them here: https://we.tl/hM5a98frZk

  • Like 1

Share this post


Link to post
Share on other sites

Thank you, I will run them later although I don't suspect the results will be much different from yours. Not sure about OpenCL either, I didn't see any load on the GPU with GPU-Z when I was testing some FLIP stuff (OpenCL enabled on the solver).

Here is the boolean sphere scene in case anyone else has trouble accessing it: https://drive.google.com/open?id=0B0BbxninBS49aEVYa21TR3JYclU

  • Like 1

Share this post


Link to post
Share on other sites
7 hours ago, Beatnutz said:

Okey, so now counting the cache time and with OpenCL set. Not a very big gain with OC, how do I know if I'm using CPU or GPU?

Grain
1950x (old) - 2:21min
1950x - 1:51min
1950x @ 3.8GHz - 1:48min

Pyro
1950x (old) - 7:37min
1950x - 6:35min
1950x @ 3.8GHz - 6:32min

Fluid
1950x (old) - 2:07min
1950x - 0:32min
1950x @ 3.8GHz - 0:30min

For OpenCL CPU you should put this in your Houdini environment file: 

HOUDINI_OCL_DEVICETYPE=CPU

Look up the manual for where to do that for Windows. Then in Houdini go to About Houdini in the Help menu and scroll down to the openCL section and it should say the Threadripper is the OCL device.

Those Fluid benchmarks look suspiciously fast, seems to be just the particles being cached not the mesh.

 

Share this post


Link to post
Share on other sites
46 minutes ago, DaJuice said:

Not sure about OpenCL either, I didn't see any load on the GPU with GPU-Z when I was testing some FLIP stuff (OpenCL enabled on the solver).

Flip OpenCL mainly improves viscosity, the pressure solve in general sees minor improvements in my testing so far.

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

×