Is AMD potentially risky? (Threadripper)

[Noobini]

hmm...needs clarifications please..

So now if I open up the Grain file, simply hit Save to Disk on the Output node, I get 1:29 which is FASTER than my first result.

[Diorn]

3 minutes ago, Noobini said:

hmm...needs clarifications please..

So now if I open up the Grain file, simply hit Save to Disk on the Output node, I get 1:29 which is FASTER than my first result.

Hmm that really is interesting, maybe someone could clarify why this is so. It would be logical that bypassing the writing to disk and just cooking to RAM would be faster... Could be that writing to disk puts the cooking in a more efficient background mode, but i need someone to confirm this...

Edited September 4, 2017 by Diorn

[Noobini]

the Grain result didn't really surprise me coz if I hit Play and watch the screen updating...THAT is the cause for the slowdown...like in Realflow...you don't wanna watch the sim updating the viewport...just write it out to disk...which is why it's faster now that I use the Save to Disk method and skip 'viewport updating'...

[Diorn]

5 minutes ago, Noobini said:

the Grain result didn't really surprise me coz if I hit Play and watch the screen updating...THAT is the cause for the slowdown...like in Realflow...you don't wanna watch the sim updating the viewport...just write it out to disk...which is why it's faster now that I use the Save to Disk method and skip 'viewport updating'...

Yes that sounds right, so for future reference let's keep using the save to disk method as we did untill now. Anyway that is a good score coming out of an older 4 core CPU. You should also try the boolean sphere benchmark, btw what speed does your ram run at?

Edited September 4, 2017 by Diorn

[Noobini]

RipjawsX F3-1600C9D-16GXM (2x8Gb)......I don't OC the RAM at all....although in bios I load the ROG OC profile which tops out at 4.7Ghz for CPU....I toned it down a bit to 4.6Ghz but didn't tinker with the RAM so not exactly sure what speed it's at...

[Diorn]

6 minutes ago, Noobini said:

RipjawsX F3-1600C9D-16GXM (2x8Gb)......I don't OC the RAM at all....although in bios I load the ROG OC profile which tops out at 4.7Ghz for CPU....I toned it down a bit to 4.6Ghz but didn't tinker with the RAM so not exactly sure what speed it's at...

Looks like those run @ 1600 mhz, but u can check those with CPU-Z software utility

[Noobini]

yeah sure can tomorrow when i should try the longer pyro/bool tests

[Diorn]

17 minutes ago, Noobini said:

yeah sure can tomorrow when i should try the longer pyro/bool tests

Cool, also if you have any ideas for other kinds of benchmarks scenes, more SOPS intensive, such as attribute transfer, for each loops, or any other things you saw are intensive from experience, plese don't hold to make one that we can add to this thread.  

[Guest tar]

If possible I would like to see TR tested with limited cores via the -j12 -j8 and -j4 settings. This would show how well Houdini linearly scales with these massive core chips. Secondly, testing on Linux vs Windows would be good too.

[einhander]

I just today finished my Treadripper build.  I use renderman and the TR overclocked to 3.9ghz renders only 10% better than my 4-core i3770k overclocked to 4.6ghz. This is using the same maya scene. If I set the buckets to 64 on each machine then the 4 core actually beats the 16 core.     I get great stat smashing results on CB and cpuid. TR in renderman acts like it's window shopping. 32 buckets doing very little. I can only assume it's due to the Embree. Any help would be appreciated.

[Noobini]

here's my latest results

Win7, i7 4770K @ 4.6GHz, 16G RAM DDR3 1600, GTX 1070

Grain - 1:33 (Save to Disk, read result from Perf. Monitor)
Fluid - 0:22 (Save to Disk, read result from Perf. Monitor) (so where do I do the 'meshing' if this seems suspiciously fast ?)
Pyro - 10:02 (Save to Disk, read result from Perf. Monitor)
Boolean Sphere - 42:42 (Render to Disk, read result from Render Scheduler Elapsed)

(OC:
1 core  @ 4.6
2 cores @ 4.4
3 cores @ 4.2
4 cores @ 4.0)

(RAM uses XMP1.2 profile which means 800MHz instead of 666...yes confusing from 1600MHz..I think advertisers would like to advertise the D(ouble)DR number rather than 800)

(Save to Disk: Samsung 850 EVO 250GB)

Edited September 5, 2017 by Noobini

[Guest tar]

28 minutes ago, einhander said:

I can only assume it's due to the Embree

Why do you think that?  A simple Google "Embree is optimized for Intel CPUs supporting SSE, AVX, AVX2, and AVX-512 instructions, and requires at least a CPU with support for SSE2." which the TR should have and your i3 has, (edit) except the super new AVX-512

Edited September 5, 2017 by tar

[einhander]

You are right. TR has all those instructions.  I am at a loss. Every other application I use works fine.

[Diorn]

7 hours ago, Noobini said:

here's my latest results

Win7, i7 4770K @ 4.6GHz, 16G RAM DDR3 1600, GTX 1070

Grain - 1:33 (Save to Disk, read result from Perf. Monitor)
Fluid - 0:22 (Save to Disk, read result from Perf. Monitor) (so where do I do the 'meshing' if this seems suspiciously fast ?)
Pyro - 10:02 (Save to Disk, read result from Perf. Monitor)
Boolean Sphere - 42:42 (Render to Disk, read result from Render Scheduler Elapsed)

(OC:
1 core  @ 4.6
2 cores @ 4.4
3 cores @ 4.2
4 cores @ 4.0)

(RAM uses XMP1.2 profile which means 800MHz instead of 666...yes confusing from 1600MHz..I think advertisers would like to advertise the D(ouble)DR number rather than 800)

(Save to Disk: Samsung 850 EVO 250GB)

Yeah probably all these benchmark scenes could use an improvement of approach, the boolean scene seems reliable as a performance result but it is unclear wich kind of opperations benefit from what kind of resource without more isolated operation testing. About the RAM, if you are reading it from CPU-Z you are supposed to double the reading  ram is dual data rate so 800x2 = 1600

[Guest tar]

10 hours ago, Diorn said:

 but it is unclear wich kind of opperations benefit from what kind of resource without more isolated operation testing. 

This can be more complicated than it seems to the user - The devs on the SideFx forum write that multithreading grain sizes and core counts affect different use cases within Houdini.  

Edited September 5, 2017 by tar

[Diorn]

Ok guys, got all the parts today and took me a few hours to assemble. Spoiler: 7900x still gets it's ass kicked in multithread cinebench: 2371 points (but i knew that before buying it) the rest of the benchmarks get more interesting. I'll post an update tomorrow...

[Diorn]

So here are the results for i9 7900x running all 10 cores @ 4.4ghz, RAM@3000mhz

Grain
1:31 - 17.08% faster than TR [1:48]

Pyro
6:03 - 7.6% faster than TR [6:32]

Fluid
0:26 - 14.2% faster than TR [0:30]

(TRscores from Beatnutz  1950x @ 3.8GHz,   RAM @ 2133MHz)

 

Boolean Sphere score:

15:43 

01.05  %   faster than TR 1950x (16 cores)(DaJuice's 1950x @ 3.75GHz, RAM @ 3200MHz)   [15:53]

12.02  %   faster than TR 1950x (16 cores)(Beatnutz  1950x @ 3.8GHz,   RAM @ 2133MHz)  [17:46 ]

25.26  %   faster than Ryzen 1700 (8 cores) (Marty's score in performance mode  Ryzen 1700 @3.68GHz) [20:20]

 

 

Cinebench scores
Multithread:2371  28.65% slower than TR (Beatnutz  1950x @ 3.8GHz,   RAM @ 2133MHz)

Singlethread:188  30.67%  faster than TR (Beatnutz  1950x @ 3.8GHz,   RAM @ 2133MHz)

 

 

Still not totally convinced about the fluid benchmark, it goes too fast for everyone, i think it needs more resolution and also meshing involved, it well may be that the test is so short that the CPU does not even get enough time to distribute the tasks to all resources. Not totally convinced about the Boolean sphere because it is a mostly CPU mantra rendering test. But i think the i9 does it's job well for what i needed it to, a general task performer in Houdini, excluding any CPU rendering needs. I'll see if i can dig up some intensive sop projects i did a short while ago, took an RBD sim of pieces blowing away, then used a for each loop to scale them to 0 and blast as they get older.

4.4 ghz on the i9 was kind of out of the box, did not change any settings manually in the BIOS yet, just turned on an XMP profile. I'm cooling it with a 360 AIO watercooling solution (closed loop, just install it and run it).

Would be cool to see more scores from the 3200mhz RAM TR build (for the 3 sims benchmarks).

 

i added the results from this thread to a spreadsheet it can be found here: https://docs.google.com/spreadsheets/d/1jQFxPdOBcWbYtJJkVafr4VYojhkV5MqInZmEsT_zmZA/edit?usp=sharing

Please let me know if you noticed i missed anything or put the wrong numbers somewhere. 

 

 

 

 

Edited September 6, 2017 by Diorn

[Guest tar]

7900x looks great! Thx for testing.

9 minutes ago, Diorn said:

66.85  %   faster than Ryzen 1700 (8 cores) (Marty's score in performance mode  Ryzen 1700 @3.6GHz) [31:30]

 

This should be '20min 20sec' when at @3.68Ghz. cheers

[Diorn]

5 minutes ago, marty said:

7900x looks great! Thx for testing.

This should be '20min 20sec' when at @3.68Ghz. cheers

Thanks, i've edited it!

[Guest tar]

@Diorn it would be great to see how OpenCL CPU performs on the 7900x too.