Is AMD potentially risky? (Threadripper)

[Diorn]

Small update, tested my workstation from the office with I7 6950x @4k for all 10 cores, cached to SSD, ram@ 2333 mhz

Boolean sphere scene: 18:07

[Beatnutz]

OpenCL CPU gave me pretty bad results. I'm trying lower ram speeds as you can see because Cinebench for some reason gives me way better scores when my ram is at 2133MHz. Houdini works better when it is set to 3000MHz. Strange.


Grain
1950x @ 3.8GHz, DDR4 @ 3000MHz - 1:48min
1950x @ 3.8GHz, DDR4 @ 2133MHz - 2:00min
1950x @ 3.8GHz, DDR4 @ 2133MHz (OpenCL CPU) - 2:31min

Pyro
1950x @ 3.8GHz, DDR4 @ 3000MHz - 6:32min
1950x @ 3.8GHz, DDR4 @ 2133MHz (OpenCL CPU) - 10:26min

Fluid
1950x @ 3.8GHz, DDR4 @ 3000MHz - 0:30min
1950x @ 3.8GHz, DDR4 @ 2133MHz (OpenCL CPU) - 0:34min


Cinebench scores
1950x @ 3.4GHz, 2133Mhz ram - 2931pt
1950x @ 3.8GHz, 2133Mhz ram - 3164pt
1950x @ 3.8GHz, 2800Mhz ram (manual) - 2939pt
1950x @ 3.8GHz, 3000Mhz ram (XMP2) - 2972pt
1950x @ 3.8GHz, 3066Mhz ram (XMP1) - 2948pt

 

Edited September 1, 2017 by Beatnutz

[Diorn]

Updates on the i7 6950x as follows for the 3 sim benchmark collection:

Grain
6950x @ 4GHz, DDR4 @ 2399MHz - 1:47min    0.9% faster than TR

Pyro
6950x @ 4GHz, DDR4 @ 2399MHz - 7:08 min      8.78% slower than TR

Fluid
6950x @ 4GHz, DDR4 @ 2399MHz - 0:34min    12.5% slower than TR

(compared to Beatnutz Threadripper scores with 1950x @ 3.8GHz, DDR4 @ 3000MHz )

 

And for tidyness i'll add again the Boolean Sphere score:

6950x (10 cores)@ 4GHz, DDR4 @ 2399MHz - 18:07min  

13.13%   slower than TR 1950x (16 cores)(DaJuice's 1950x @ 3.75GHz, RAM @ 3200MHz) )   

1.9%      slower than TR 1950x (16 cores)(Beatnutz  1950x @ 3.8GHz,   RAM @ 2133MHz)

11.53%   faster than Ryzen 1700 (8 cores) (Marty's score in performance mode  Ryzen 1700x @3.6GHz)

30.5882% faster than dual Xeon 5680. Total 12 cores at 3.33MHZ. 24min 30sec

 

 

 

Cinebench scores
Multithread

6950x @ 4GHz, ram @ 2399MHz - 1950    

47.47% slower than TR (1950x @ 3.8GHz, 2133Mhz ram ) =  3164

48.51% slower than TR  (1950x @ 3.75GHz, 3000Mhz ram) = 3199 CB score

Singlethread

6950x @ 4GHz, DDR4 @ 2399MHz - 168  19.6%  faster than TR (1950x @ 3.8GHz, DDR4 @ 2133MHz)

 

 

Thoughts:

Have to say the I7 stacks up rather good against the TR for sims and general H use, at least from these benchmarks, and sadly i have slower RAM in this build. But then again it was a 1750$ CPU. Looking forward to try out the i9 as it can easily do 4.4 ghz.

Clearly TR is a beast for CPU Rendering, personally i am reasured that i won't be doing any of that with the i9 - (But Cinebench is the only test here, would be cool to make a test scene for Mantra, and something separate more SOPs involved. For each loops and attribute transfer are not fully multithreaded for example). Right now it seems the cinebench score is almost to the dot exactly the percentage betweet the numbers 10 vs 16 being 46.15% (mind you it is not the latest intel, it is overclocked mildly and the rams differ).

I see tasks not involved with CPU rendering benefiting a lot to high CPU speed with high speed ram, with ram being the most interesting find. (Old PC building logic for performance stays true, faster but less cores might beat more cores@slower speeds unless we are talking about highly specialized and properly ported multithreading tasks). Either way you go, make sure to get at least 3000mhz ram, maybe even the CAS latency matters (ex:cl14 vs cl16) but have no way of knowing this for real at the moment, unless we start to mention this as well in the scores.

 

PS: @Beatnutz thanks for trying out different ram speeds, it looks like they really do count. Could you also add single thread scores for Cinebench please?

Edited September 4, 2017 by Diorn
added CB scores, added Boolean sphere scores and Ram speeds

[kemijo]

Great to see these benches! Beatnutz (love the name, if it refers to the rap duo), that is very interesting regarding the ram speeds. Wonder if that is specific to Ryzen/TR.

Correct me if I'm wrong but the results might be more indicative of performance if they were more intensive (i.e. longer times) like the boolean sphere test. I think 0:30, 1 min and 2 min tests are too close to the 'noise floor' so to speak, to be reliable. Yes, no?

Anyone with a recent i7 (6850 or so) that can run these tests too? Would be great to see old i7, new i7, i9 and Threadripper stress tests specifically in Houdini.

[Diorn]

15 minutes ago, kemijo said:

Great to see these benches! Beatnutz (love the name, if it refers to the rap duo), that is very interesting regarding the ram speeds. Wonder if that is specific to Ryzen/TR.

Correct me if I'm wrong but the results might be more indicative of performance if they were more intensive (i.e. longer times) like the boolean sphere test. I think 0:30, 1 min and 2 min tests are too close to the 'noise floor' so to speak, to be reliable. Yes, no?

Anyone with a recent i7 (6850 or so) that can run these tests too? Would be great to see old i7, new i7, i9 and Threadripper stress tests specifically in Houdini.

Agree about longer performance tests, i also add we should have one for Bullet and some more SOP intensive ones with attribute transfer and for each loops. I can run them at work with a 10 core i7 and at home with the 10 core i9 when i get all the parts. For giggles i could also do my current i7 4770k

Edited September 1, 2017 by Diorn

[Beatnutz]

Feel free to make a new HIP we can test. Preferably not too long though because I'm already benchmarking 24/7 here instead of working

[Beatnutz]

@Diorn

I think you missed your Cinebench scores?

[Diorn]

3 minutes ago, Beatnutz said:

@Diorn

I think you missed your Cinebench scores?

added them,

hah tell me about it, i'm swamped with work as it is

Edited September 1, 2017 by Diorn

[Beatnutz]

15 hours ago, Diorn said:

 

PS: @Beatnutz thanks for trying out different ram speeds, it looks like they really do count. Could you also add single thread scores for Cinebench please?

Cinebench, single core
1950x @ 3.8GHz, DDR4 @ 2133MHz - 138

Could probably beef that number up by using 'game mode' but it's a pain in the butt to activate and a pretty useless mode overall. Don't think a lot of TR users will use it.

Edited September 1, 2017 by Beatnutz

[Diorn]

2 minutes ago, Beatnutz said:

Cinebench, single core
1950x @ 3.8GHz, DDR4 @ 2133MHz - 138

Thanks! updated my post with comparisons

[Beatnutz]

BooleanSphere
1950x @ 3.8GHz, DDR4 @ 2133MHz - 17:46min

RAM frequency obviously does a lot on this one

Edited September 1, 2017 by Beatnutz

[Thomas Helzle]

Interesting thread, thanks everybody!

Reading about those 18 core TRs or new intel cpus made my mouth water initially, but what I read here doesn't exactly make me want to jump in right away. My 6 core i7 3930 from 2013 overclocked to 4.1 Ghz (costing about 500 Euro back then) with 64 GB GDR 3 RAM @1600 takes 30 minutes for the Boolean Sphere scene - not too shabby. So while you guys get almost twice the speed, that still isn't exactly as earth shattering as I would have expected.

Cinebench Single Core is 139 here, so basically nothing much has changed in that department. So it's only really worthwhile for CPU rendering (I only score 1034 in Cinebench with 6 cores / 12 threads) but I use the GPU for that anyway.

I think I'll wait some more with getting a new machine and let things settle down a bit regarding drivers, prices and bioses... :-)

Cheers,

Tom

Edited September 2, 2017 by Thomas Helzle

[Guest tar]

Those OpenCL CPU results are a worry. Seems like Intel OpenCL CPU drivers aren't optimised for 32 threads or something more nefarious . I'm sure SideFx are aware of it but it's probably worth sending in a bug report to them. Thx!

 

also ran the BooleanSphere on the Mac dual Xeon 5680. Total 12 cores at 3.33MHZ. 24min 30sec.

 

@Diorn it's a normal 1700 not the 1700x. 

Edited September 3, 2017 by tar

[Diorn]

On 9/3/2017 at 8:40 AM, marty said:

Those OpenCL CPU results are a worry. Seems like Intel OpenCL CPU drivers aren't optimised for 32 threads or something more nefarious . I'm sure SideFx are aware of it but it's probably worth sending in a bug report to them. Thx!

 

also ran the BooleanSphere on the Mac dual Xeon 5680. Total 12 cores at 3.33MHZ. 24min 30sec.

 

@Diorn it's a normal 1700 not the 1700x. 

added the Xeon score, and edited the ryzen, thanks!

Guys, this thread on the facebook HoudiniArtists group is also a good collection of benchmark scores. It uses a scene that does volumes and VDB opperations

https://www.facebook.com/groups/HoudiniArtists/search/?query=benchmark

 

Edited September 4, 2017 by Diorn

[Noobini]

what's the procedures to benchmark in Houdini ? is it:

01) open scene

02) Performance Monitor, Record

03) type in End Frame number in Current frame box, press Enter (or do you just hit Play here and wait for it to reach the end?)

04) Stop Record when finished, read Total time

(just thought I'd run my crusty old rig thru the tests for fun)

Edited September 4, 2017 by Noobini

[Diorn]

1 hour ago, Noobini said:

what's the procedures to benchmark in Houdini ? is it:

01) open scene

02) Performance Monitor, Record

03) type in End Frame number in Current frame box, press Enter (or do you just hit Play here and wait for it to reach the end?)

04) Stop Record when finished, read Total time

Good question, this is the workflow i did yes but for 3) i would say that your option would bypass any disk cacheing.

I think we were all doing the write to disk at the final node up untill now.

 

Either way use the performance monitor, when it is armed the timer starts and stops togheter with the start of the cacheing and end of the cook, you just need to de-arm it at the end.

 

Edited September 4, 2017 by Diorn

[Noobini]

ok, here's my results on an old OC rig...non-cheat method, ie. hit Play button instead of typing in end frame/Enter

i7 4770K @4.6Ghz, GTX 1070, 16G DDR3

Grain - 1:48 (1:24 if cheating)
Fluid - 1:17 (0:20 if cheating !!!)

Pyro..tomorrow

(@4.6Ghz that's 1 core, 2 cores @4.5, 3 cores @4.4, 4 cores @4.3)

 

 

Edited September 4, 2017 by Noobini

[Diorn]

2 minutes ago, Noobini said:

ok, here's my results on an old OC rig...non-cheat method, ie. hit Play button instead of typing in end frame/Enter

i7 4770K @4.6Ghz, GTX 1070, 16G DDR3

Grain - 1:48
Fluid - 1:17

Pyro..tomorrow

(@4.6Ghz that's 1 core, 2 cores @4.5, 3 cores @4.4, 4 cores @4.3)

 

 

Nice OC-ing there! have the same 4 core I7 home, running it at a more conservative @4.2 ghz. I think you mean core 1,2,3 and 4 because at first i was counting 10 cores

But already we are adding inconsitencies with the results as i am sure everyone used the cache write files to disk for the scores not just hitting play. 

 

[Noobini]

don't I just Record, then hit the Up arrow and watch the Timeslider slowly creep across the screen ?

(the OC at top speed @4.6 applies when only a single core is needed..so it can run at top speed, when all 4 cores are needed...it cannot sustain all 4 cores at 4.6...so the more cores running concurrently, the slower the OC is..ie 4 cores @4.3)

Edited September 4, 2017 by Noobini

[Diorn]

3 minutes ago, Noobini said:

don't I just Record, then hit the Up arrow and watch the Timeslider slowly creep across the screen ?

your option would bypass any caching to disk.

I think we were all doing the write to disk at the final node up untill now. waiting for the guys to confirm. Just hitting play without writing to disk would give you smaller timings (cheating, but it is isolating the CPU and RAM more)

Edited September 4, 2017 by Diorn