Choosing the most efficient HW for RC

Answered

Comments

7 comments

  • Avatar
    chris

    looks pretty good.

    and as long as the models don't get too big. it should be fast.

    though 512gb 960 pro for temp drive, and 64gb of ram are probably border line but should be fine 95% of the time esp if your good and clearing temp drive between runs.

    I find with 128gb ram that i'm never running out with 2500, but i do some of the time if i'm running cli.

    temp drive i regularly use more than 512gb.

    I also find the temp drive for rc can't use more than about 35-45 mb/s. but a hard drive will be really slow. I think it more about response time vs transfer speed. but i guess your looking at QD1 speeds, which is about 45mb/s for a 960. I've run this on quite a few variations, from hard drives (bad don't do it). single 960s, 2 x 950 in raid 0 and 4 x sandisk ssds in raid 0. Once you start using at least a normal ssd there isn't much of an improvement beyond that. and size becomes more important.

    though if you do run out of ram and it goes to swap then the 960's, and raid 0 ssd's arrays get much but better speeds there.

     

     

    0
    Comment actions Permalink
  • Avatar
    Jennifer Cross

    cpu - Dual xeon wil help you more the the i9 in most of the cpu based computation sections.  There are a few places where it single thread only, but most of the time it keeps all 16 cores in my box (see below) busy.  Find some older server gear with enough pci-e x16 slots and you should get a good platform to build on.

    Also you may want to consider just renting cloud time rather than investing in hardware that you only work hard a few hours a week.  AWS has some nice instances for the (p3) which scale for power (and cost) to some really impressive configurations (p3.16xlarge - 64 intel core, 430gb ram, 8x tesla v100 gpu etc)

    https://aws.amazon.com/ec2/instance-types/p3/

    0
    Comment actions Permalink
  • Avatar
    chris

    I'm interested in how the dual xeons go. is rc ok with multiple numa nodes?

    have you run the benchmark on your xeons?

    I should also do that too soon when i get a chance.

    would also be nice if someone tested with ryzen, threadripper, and epyc

    0
    Comment actions Permalink
  • Avatar
    Jennifer Cross

    Hi Chris

    I'm not sure what/which benchmark you mean but RC seems to run fine - there are a few places where it drops down to single threaded, but mostly it is pretty good at keeping the cpu/gpu pipelines fed. (doesn't use a lot of gpu memory for processing - just a hog when it wants to display the mesh in "sweet").  I've had it use about 40gb ram on some models but it does do a lot of disk paging to the temp area - even with a bunch memory free.

    This is where the new systems could be useful (cpu attached NVMe drives in raid 0 for swap/temp space)

     

    But the dell seems to be fast enough for general work

    0
    Comment actions Permalink
  • Avatar
    chris

    https://support.capturingreality.com/hc/en-us/community/posts/115001227911-Hardware-Optimisation-Benchmarking-Shenanigans-

    have in look in that thread.

    I'd be interested in how the dual xeon goes with model generating part of the process, which is the slowest part for me.

     

    0
    Comment actions Permalink
  • Avatar
    Benjamin FONTAINE

    Hi everyone ;)

    I discovered this program one week ago and I'm very enthusiastic with using it ;)

    => I'm hesitating to replace my ram for 124 bucks in order to achieve 32go

    But I'm somehow I can but 1T M2 2280 SSD for 150 bucks .. Thing is how fast might a swap SSD behave ?

    Since I'm willing to test a rendering with more than 1500 pictures I don't thing 32g would be enough .. is this relevant or not at all ?

    Do you have any clues, notes, comments, enlightement on this case ?

    Regards, Ben

    0
    Comment actions Permalink
  • Avatar
    Lucia CR

    hello Benjamin,

    using an SSD is our recommendation, the memory consumption depends on the number of images (irrespective of their size) and the number of detected features per image, 32GB RAM should be ok even for 4,000 images (processing with default settings)

    0
    Comment actions Permalink

Please sign in to leave a comment.