CUDA Compute and PCIe Lanes

Discussion in 'Machine Learning, Deep Learning, and AI' started by Jeggs101, May 12, 2017.

  1. Jeggs101

    Jeggs101 Well-Known Member

    Joined:
    Dec 29, 2010
    Messages:
    1,466
    Likes Received:
    216
    It appears as though we may have some budget for I was considering going Ryzen for my CUDA build. Since it seems like CUDA is the new STH trend.

    Here's what I'm mulling for work. I'm going to start trying dl and my boss wants to do the same. My plan is to install two GPUs, one for each of us.

    On Ryzen that means each GPU gets pcie 3.0 x8 only if we use a x370 mobo.

    Does x8 really matter? For gaming I don't think it does.
     
    #1
  2. _alex

    _alex Active Member

    Joined:
    Jan 28, 2016
    Messages:
    874
    Likes Received:
    94
    following this, just did some research but only found benchmarks measuring fps.
    there pcie 3.0 x8 vs x16 seems to be <1%

    i consider putting one or two 1070 on a s2600 2011 system with Dual e5 v1 CPU.
    this would unfortunately mean pcie 2.0 x8 for the GPUs or upgrading to V2 CPUs.

    on the gaming-sites quite a few people think 3.0 x4 would still be enough for a 1080, what is in the range of 2.0 x8 bandwith.

    any opinons (or even numbers) if and how much pcie 2.0 / 3.0 x8 would affect machine learning ?

    trying to estimate if it's worth looking for V2 CPU or getting cheaper GPU that can't saturate pcie 2.0 x8 if a 1070 would be bottlenecked in a way that it's just a waste of money for this environment.
     
    #2
  3. Jeggs101

    Jeggs101 Well-Known Member

    Joined:
    Dec 29, 2010
    Messages:
    1,466
    Likes Received:
    216
    @_alex thanks! I didn't make a decision yet.

    Doesn't 3.0 also give better latency?
     
    #3
  4. _alex

    _alex Active Member

    Joined:
    Jan 28, 2016
    Messages:
    874
    Likes Received:
    94
    in theory, it should.
    but this would really depend if the application feeding the GPU is able to reach any limit theoretically introduced by fewer and/or slower pcie lanes.

    Sorry, can't say anything more meaningful. I basically have the same question like you, just extended how bad pcie 2.0 x8 would be for GPU-compute.
     
    #4
  5. LukeP

    LukeP Member

    Joined:
    Feb 12, 2017
    Messages:
    156
    Likes Received:
    14
    never look at fps. its entirely irrelevant. the result is entirely dependent on the algorithms used. if you are Model distributed you need more bandwidth than data distributed.

    if you dont know what im referring to, it probably wont matter to you. just get one fast gpu. and relax.
     
    #5
    KioskAdmin likes this.
Similar Threads: CUDA Compute
Forum Title Date
Machine Learning, Deep Learning, and AI Learning cuda from scratch May 21, 2017

Share This Page