Question about GPU Utilization and CPU Utilization during ML

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

T_Minus

Build. Break. Fix. Repeat
Feb 15, 2015
7,640
2,058
113
Would it be fair to think that if you're currently utilizing an aws instance of p2.xlarge (4 vCPU 1x K80 GPU 64GB RAM) and not nearly maxing it out that you could go to your own dedicated server with 2 K80s (or what made most $ sense) and 2x 10C CPU and have enough CPU resources to run VMs and do other tasks without affecting the machine learning VMs utilizing the K80s?

p2.xlarge from amazon is not cheap, and I have a spare GPU chassis/motherboard as well as everything except the GPUs and am wondering if I can help a friend out not only on the GPU side but allow him to do a lot more with the more cores and ram (256GB RAM) than he is now.
 
  • Like
Reactions: Tha_14

Legotan

New Member
Jul 27, 2018
18
2
3
I think it should be possible. I mean if you host your "own" p2.xlarge VM you still got 192GB RAM and 16-36 Cores (nonHT/HT) for your other VMs.
 

Patrick

Administrator
Staff member
Dec 21, 2010
12,513
5,804
113
Yes. BTW if you want K80 systems, I know a place with lots of Dell 4x K80 systems, dual E5-2680 V3's, 2TB P3700's, and 128GB RAM (you would want more) in the $7K each range.
 
  • Like
Reactions: T_Minus

T_Minus

Build. Break. Fix. Repeat
Feb 15, 2015
7,640
2,058
113
Thanks guys.

Right now I'm sitting on the following systems that I can convert to GPU based servers.

- SM X9DRX + SM Chassis (4U/Tower/Conversion)
- SM X9DRX (No Chassis at this time, maybe keep as a spare)
- SM X10DRX + SM Chassis (3U w/GPU Power Cables)

-- I already have CPUs for the X10DRX (14 Core V3).
-- No CPU yet for the V2 systems seems around $600 would get me 2x of the higher-end V2s
(I do have lower core/frequency V2s)


Questions:
  • Are the higher-end V2 detrimental to the GPU utilization compared to V3s?
  • What about DDR3 vs DDR4?

    Not sure what kind of performance hit I'd take on the x9s?

I have enough spare parts to put together the X10 system 100% minus GPU, so that one is a no-brainer, I'll use it.
I have everything excluding the CPUs (I do have 2 trays of CPUs but they're lower-core/frequency so I could use them to 'start').

Has anyone run the same GPUs for ML and compared from a v2 with ddr3 to a v3 with ddr4?

I have a hunch that starting out the v2 systems will be sufficient, I also believe I can get by with 1 of these for awhile since I can put a few GPUs in them... so maybe sell the two x9s and hold out for an x11 in the future and use that $ to buy a GPU to start out?

Thoughts?