Author Topic: XEON 2699 V3 bad scalability  (Read 7341 times)

2015-12-02, 21:56:35

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
Hello everyone, I'm doing a test with this great render engine but one thing puzzles me a lot, doing test render with the xeon 2699v3 (18 cores 36 threads) when using 20 cores the same scene same settings time is 15 min, with 36 cores 19 minutes, is normal so little difference ???? I have two procesors like this such dual mount plate, but the scalability seems bad, I did not like the height of the processors that are ...they work best in each separate network processors or dual mode ??

5 Cores: 1h 28 min.
8 cores: 54 min.
10 cores:45 min.
12 cores: 38 min.
16 cores: 29min.
36 : 19 min.

Greetings and sorry for my bad English.
« Last Edit: 2015-12-02, 22:18:43 by Txamo »

2015-12-02, 22:57:05
Reply #1

Juraj

  • Active Users
  • **
  • Posts: 4815
    • View Profile
    • studio website
Few notes:

Your post is confusing because you keep mixing threads and cores. Pick only one and don't mention the other one in equation. Let's go with Cores, because it's better for pure scalability, ignoring hyper-threading contribution.

-Are you comparing different CPUs ? Or single CPU with setting different thread allowance in Corona settings ? I presume the latter, because there is no 5 or 10 core CPU.
-Always multiply frequency by core amount to get your total computational power. In theory, 8 cores @ 4 GHz will be equal to 16 cores @ 2 Ghz. In practice, the first will be slightly stronger because nothing is perfectly multi-threaded.

-Corona scales well in latest (1.3) version, are you using this one for you testing ?
Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2015-12-02, 23:13:14
Reply #2

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
If Juraj, you're right I explained wrong. I'm talking about a single processor  with 18 cores, though I have not installed the two processors in the dual plate, these tests are in an X99 board from another team. I'm comparing the same processor but with different active cores in the corona render settings.
I am using corona render 1.3.

If I understand correctly what you say Juraj, this 18-core Xeon 2.3 Ghz give a total computing power of 4.1 and my other workstation i7 5960x 3 GHz with 8 cores give a total of 2.4 is nearly double power, but this translates into half time ?? I could not even testing but makes so that everyone knows. If I put the two Xeon dual plate up out party power is better to have them both or separately?

Thank you very much and sorry I'm a little green in some things.

2015-12-02, 23:22:47
Reply #3

Juraj

  • Active Users
  • **
  • Posts: 4815
    • View Profile
    • studio website
Single Xeon E5 2699 v3  18@ 2.3Ghz =  41,4 Ghz
i7 5960X  8x3.5Ghz =  28 Ghz   (note: I am using turbo-frequency, but most users overclock into 4.0-4.2 Ghz manually)

So single 2699 will be 35 perc. faster in theory, and slightly less in practice, because not everything is perfectly multi-threaded.

Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2015-12-02, 23:29:35
Reply #4

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
And working on a dual motherboard this two xeons scalability will be proportional? not if I'm wrong but seeing this does not seem worthwhile or perhaps better put Xeones working on two motherboards as perhaps render node ??

2015-12-02, 23:53:45
Reply #5

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
Maybe better idea to put each xeon separately in different motherboards to get more power of each ?? I doubt give their full potential in a dual motherboard.

2015-12-02, 23:56:31
Reply #6

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 9048
  • Turning coffee to features since 2009
    • View Profile
So single 2699 will be 35 perc. faster in theory, and slightly less in practice, because not everything is perfectly multi-threaded.

Sidenote: rendering with PT+PT or PT+UHD where UHD is already computed actually is almost perfectly multi-threaded (lets say 99.99%). What gets you is the memory - you add CPU cores, but you dont add independent memory controllers with memory chips. Single memory controller/chip cannot serve 20 cores as efficiently as a single core.
Rendering is magic.How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2015-12-03, 00:03:57
Reply #7

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
forgive Ondra, me lose a little technicalities and do not quite understand the point you've made, it's better to take full power of the Xeon 2699, put them on separate motherboards or in a dual motherboard ?? (I have to ask the board questions and now I come)

Thank you very much to you both for the help.

2015-12-03, 01:00:17
Reply #8

Juraj

  • Active Users
  • **
  • Posts: 4815
    • View Profile
    • studio website
It's going to be 'almost' proportional. (dual-socketing). These CPUs are so expensive for exactly that single reason :- )

The loss will be much less significant than having to buy double the pc parts and using distributed rendering :- )
Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2015-12-03, 01:11:19
Reply #9

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
Thanks Juraj, then will go forward with the purchase of the motherboard Asus Z10PE-D8 WS, after what I was hearing here doubted if they would not give either the xeon based on dual motherboard a so wondered. The truth is that the processors left me very cheap 2,100 euros two (I have a contact in Costa Rica processors happens to me, you already have the new 24-core xeon ... crazy). I started a thread a few months ago we talked about this Juraj but will not remember, you told me that when she was put an assembled all the results, well now when I get all the will detail, promised.

I did not understand very well the explanation of Ondra, that was what he meant.

Thank you very much for your time.

2015-12-03, 03:40:03
Reply #10

Juraj

  • Active Users
  • **
  • Posts: 4815
    • View Profile
    • studio website
Ondra simply corrected factuality that it is not multi-threading when it comes to Corona rendering, but memory access (hardware issue) that limits the potential of absolute 100perc. scaling.

So you were able to get 2699 v3 for 1050 euro each including VAT ? Whoa there.... how busy is this contact of yours :- ) ? Are these ES (Engineering samples) models by any chance ?

Asus Z10PE-D8 WS is excellent choice, just saw it live in Ondra's office week ago. Looked nice :- )
Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2015-12-03, 08:48:36
Reply #11

Rhodesy

  • Active Users
  • **
  • Posts: 558
    • View Profile
Does corona scale better than the likes of cinebench? I usually use that as a go to for comparing potential cpu upgrades. So if we can add up the cores/ghz combo and get 99% accuracy for corona speed that is a useful tip. I'm assuming we only lose a little bit of speed through the ram controller for CPUs with 10cores and below? Is it quantifiable?

2015-12-03, 13:20:39
Reply #12

Nekrobul

  • Primary Certified Instructor
  • Active Users
  • ***
  • Posts: 1028
    • View Profile
I do not know about 2699 v3. But with double procesor (24 threads) 5650 we ate having absolutley 0 problems and comparing to 12 thread 3930K clocked to 4Ghz it gives about 40% speed up. And they are of courese are much much more cheaper because you can esily get used once (for example we bought a cople fo workstations HP Z800 with double procesor X5650 for 600 eur each, and a new blade for the sever with same procesors costs us aproximatly the same)

May be it is not corona related? have you cheked it on any other renderengiene?
---------------------------------------------------------------
https://www.blackbellstudio.com/
https://www.behance.net/blackbell3d
CEO at "Blackbell Studio"

2015-12-03, 16:04:25
Reply #13

Txamo

  • Active Users
  • **
  • Posts: 13
    • View Profile
Sorry Juraj was working. As I said on the processors, the price is with VAT, customs 50 euros. So far it is trusted the person selling me these processors, memory modules, I bought several processors and I have been in contact with many others, I know a company that has even bought 13 of these processors. I have not had any problems so far in 2 or 3 days I have them at home without problems and gives me personal guarantee of six months (not sure if this makes it to all persons I suppose that). Processors are ES versions but have not so far had no problems whatsoever and the people I know either, temperatures, voltages are correct benchmark took with them about 6 months working daily and well. I am waiting for me to get the motherboard and power supply to mount, if you want to post them here on their operation or if you have questions we can do tests.

Nekrobul have not been tested with another rendering engine mount but next week if you want to do tests, the question is because I did not know if it was right time gain per core.

Thanks to all.