Optimal mid-high to high (~7k $) GPU + CPU computing station

Message boards : Number crunching : Optimal mid-high to high (~7k $) GPU + CPU computing station
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 2022329 - Posted: 8 Dec 2019, 7:25:54 UTC

Hello
I would ask some advise what are current best in performance per $ (initial cost only, electricity not counted ) metric CPU, GPU and memory (+mobo to support selected computing devices).
Approx system price could be ~7k$ multy cores for CPU and multi-GPU setups preferable (station will be used for some quantum mechanics computations with CPU-only software and molecular dynamics computations with GPU support). Also, what are current realistic memory amounts to not to go into too high price segment? Hope some perspectives to use spare time for SETI also (unfortunately, I'm not the planned owner :) )
SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 2022329 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13882
Credit: 208,696,464
RAC: 304
Australia
Message 2022331 - Posted: 8 Dec 2019, 7:47:11 UTC - in response to Message 2022329.  
Last modified: 8 Dec 2019, 7:48:38 UTC

Hello
I would ask some advise what are current best in performance per $ (initial cost only, electricity not counted ) metric CPU, GPU and memory (+mobo to support selected computing devices).
It depends very much on the software for the CPU work. If it makes use of AVX 512, then an Intel HEDT (High End Desk Top) CPU is the only option, the performance is huge.


Otherwise, given the price limitations, a Ryzen9 would be the way to go (or a Threadripper or low end Epyc system if more funds can be made available).


The amount of system memory required would once again depend on the CPU software being used- how much does it need?
If the GPU software makes use of Double Precision, then go with a desktop AMD GPU, Single Precision then Nvidia RTX 270 Super (or RTX 280Ti if funds allow).


It's going to be a juggling act between how much CPU performance is required, and how much GPU performance, and trading off between those 2 to fit within your budget.
Grant
Darwin NT
ID: 2022331 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 2022337 - Posted: 8 Dec 2019, 9:29:49 UTC - in response to Message 2022331.  

Thanks for hint, now I'm know in what side to look :)
New models appeared I was not aware of.

Hello
I would ask some advise what are current best in performance per $ (initial cost only, electricity not counted ) metric CPU, GPU and memory (+mobo to support selected computing devices).
It depends very much on the software for the CPU work. If it makes use of AVX 512, then an Intel HEDT (High End Desk Top) CPU is the only option, the performance is huge.


Otherwise, given the price limitations, a Ryzen9 would be the way to go (or a Threadripper or low end Epyc system if more funds can be made available).


The amount of system memory required would once again depend on the CPU software being used- how much does it need?
If the GPU software makes use of Double Precision, then go with a desktop AMD GPU, Single Precision then Nvidia RTX 270 Super (or RTX 280Ti if funds allow).


It's going to be a juggling act between how much CPU performance is required, and how much GPU performance, and trading off between those 2 to fit within your budget.

SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 2022337 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2022363 - Posted: 8 Dec 2019, 16:34:06 UTC

There was a recent rant about testers using software recommended by Intel to test their processors in reviews that automatically turn off standard SIMD and AVX instructions when they detect running on any processor identifying as an AMD processor. Bench results show horrendous AVX results for any AMD processor simply because they don't even run the code on their AVX pipeline and only run it through the slowest SIMD pipeline capability. Matlab was the example software benchmark program in the review suite. It shows the new AMD processors falling way behind the Intel processor in the test suite when in fact, the AMD processor beats the Intel processor handily when run properly through the AVX2 pipeline.

Typical Intel test and marketing shenanigans against AMD that have never stopped.

So if you see any test review using Matlab, just throw out the result as it is invalid.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2022363 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2022367 - Posted: 8 Dec 2019, 16:56:45 UTC - in response to Message 2022363.  

I've heard of this issue with Matlab, but I was under the impression that it was a Matlab thing where they are using old methodology based on the old AMD chips which didnt do AVX well. I think you can edit/hack some config files on Matlab to re-enable AVX on AMD chips.

but back to the topic at hand, I fully agree with the points Grant made. Figure out the requirements of the intended software (including the operating environment; Windows/Linux/etc), and then go forward from there. Nvidia has a clear advantage on SETI/Linux, but that wont necessarily be the case for other uses.

As an example on SETI. My fastest system, and the fastest system on the whole project, doing ~1.2-1.3 Million RAC, cost less than $7000 to build, and was built at a time when GPU prices were higher. But it wont necessarily be the best bang for buck on other projects, especially ones that need double precision where it would be destroyed by a lesser AMD based system using cheap/old cards lol.

but I don't do any CPU processing on SETI. I like the c602 platform because E5-2600v2 processors are cheap, registered ECC DDR3 ram is CHEAP, and the CPUs give reasonable performance. They will do AVX, but not AVX2 or AVX-512. moving up to the newer gen systems will require DDR4 memory, which can be a factor depending how much you need. DDR4 is at least 2x the price of the reg-ECC DDR3 stuff.

3rd gen TR, or Epyc Rome are very attractive for raw CPU performance, but it comes at a huge price for the platform. expensive memory and very expensive motherboards.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2022367 · Report as offensive
Profile Zalster Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 27 May 99
Posts: 5517
Credit: 528,817,460
RAC: 242
United States
Message 2022372 - Posted: 8 Dec 2019, 17:39:20 UTC - in response to Message 2022329.  

Hello
I would ask some advise what are current best in performance per $ (initial cost only, electricity not counted ) metric CPU, GPU and memory (+mobo to support selected computing devices).
Approx system price could be ~7k$ multy cores for CPU and multi-GPU setups preferable (station will be used for some quantum mechanics computations with CPU-only software and molecular dynamics computations with GPU support). Also, what are current realistic memory amounts to not to go into too high price segment? Hope some perspectives to use spare time for SETI also (unfortunately, I'm not the planned owner :) )


I'm more interest in this CPU-only testing with quantum mechanics. Will this be a multi-thread running software like GPUGrid? Over there they were testing work units that required 4 threads for each work unit. Back when I still had my 20 core, I was finding that I was limited to 3 work units or 12 active threads due to the high temperatures I was getting running a 280mm radiator on the machine. After talking with Keith, I moved to a 420mm radiator to help keep the temperatures reasonable. The most I ever got running was 4 work units (16 threads). At the time I was running 64 GB @ 3348hz memory. That machine had 4 1080Ti on it but after I dedicated it to GPUGrid, i pulled 2 of the 4 cards out. That was an ASUS X99ew-s boardwith with the PCI3 slots running at 16X due to PLX chips. ASUS Intel WS X299 SAGE is probably the most current version for this board and you would need a chip that has at least 44 pcie lanes. You might be able to get 3600-3800 speed on the DDR4 ram. I also was using a Titanium EVGA 1600 PSU. I have no experience on AMD chips. I'll leave the discussion on GPU for you to decide on.
ID: 2022372 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2022373 - Posted: 8 Dec 2019, 17:45:49 UTC - in response to Message 2022367.  
Last modified: 8 Dec 2019, 17:47:45 UTC

I've heard of this issue with Matlab, but I was under the impression that it was a Matlab thing where they are using old methodology based on the old AMD chips which didnt do AVX well. I think you can edit/hack some config files on Matlab to re-enable AVX on AMD chips.

https://www.extremetech.com/computing/302650-how-to-bypass-matlab-cripple-amd-ryzen-threadripper-cpus
@echo off
set MKL_DEBUG_CPU_TYPE=5
matlab.exe


I was just pointing out you need to examine just what cpu architecture does best for the projected science application this server is going to process and watch out for outdated benchmarks as your example from Anandtech. I see Matlab still in use in review articles in their science application suites. Just wanted to throw awareness that application does not really show the true capabilities of AMD processors.

Doesn't really have anything to lay blame on Matlab other than they are using an Intel compiler that automatically cripples AMD. Any other company can do the same thing and use a crippled Intel compiler to make their application and it will not truly represent the performance of AMD.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2022373 · Report as offensive

Message boards : Number crunching : Optimal mid-high to high (~7k $) GPU + CPU computing station


 
©2025 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.