EVGA

RTX Titan for machine learning

Author
Post
Essentials Only Full Version
rhussain
FTW Member
  • Total Posts : 1679
  • Reward points : 0
  • Joined: 2009/07/17 06:58:49
  • Location: Infinite Loop
  • Status: offline
  • Ribbons : 7
2019/12/18 19:31:27 (permalink)
Hey guys, I need your help and advice regarding a major GPU upgrade I have in mind. I am starting to become more involved in AI and machine learning work and I am considering the purchase of two RTX Titan GPU's with educational discount from Nvidia. Here is my current build (https://www.modsrigs.com/detail.aspx?BuildID=40029). I am not sure if my PSU will be able to handle the load of both cards at 100% utilization for days or weeks at a time. I have the Corsair AX 860 watt PSU. My workload is mostly VRAM intensive which is why I need all 48GB of VRAM from bridging the two cards with NVLink. What are your thoughts? or suggestions? Anyone here have any direct experience configuring NVLink on the Titan RTX? I am wondering if these two cards can be unified as one computer device in the drivers.

 



#1

5 Replies Related Threads
    Sajin
    EVGA Forum Moderator
    • Total Posts : 49168
    • Reward points : 0
    • Joined: 2010/06/07 21:11:51
    • Location: Texas, USA.
    • Status: online
    • Ribbons : 199
    Re: RTX Titan for machine learning 2019/12/18 21:26:42 (permalink)
    You'll need a 1000w minimum. I'd get 1200w. Memory pooling will be dependent on the software you're running.


    See my current build here: https://forums.evga.com/My-bench-rig-unlimited-power-m3596024.aspx
    RTX 2080 Ti 4-way SLI?!?!: https://forums.evga.com/RTX-2080-Ti-4way-SLi-set-by-yours-truly-m2954756.aspx
    #2
    wmmills
    CLASSIFIED ULTRA Member
    • Total Posts : 5679
    • Reward points : 0
    • Joined: 2006/01/04 20:47:29
    • Location: New Jersey
    • Status: offline
    • Ribbons : 40
    Re: RTX Titan for machine learning 2019/12/19 05:49:45 (permalink)
    Sajin
    You'll need a 1000w minimum. I'd get 1200w. Memory pooling will be dependent on the software you're running.


    +1...id go 1200 if your going to be running two of them for efficiencys sake and some wiggle room.

    MOBO: EVGA x299 Dark, CPU: I9 10900X, RAM: Patriot Viper RGB 3600 32gb, SSD: Samsung 860 EVO 1TB, M2: Samsung 970 EVO+ 1TB, PSU: CoolerMaster M2 1500, CPU HSF: EVGA 240 CLC HSF~ P/P EK Furious Vardar, G-CARD:EVGA RTX3090 FTW3 Ultra Gaming w/Hybrid kit and Noctua IPPC 3000 P/P, CASE: LIAN-LI PC-V2010B w/ Window mod, OS: Windows 10 Pro 64bit, MON: Alienware AW3821DW

     New EVGA product? Register it NOW with this link:
    http://www.evga.com/register/default.asp?affiliatecode=4QFQRAMOII
     Help Our Vets From K-2!
    [link=https://strongholdfreedom
    #3
    kevinc313
    CLASSIFIED ULTRA Member
    • Total Posts : 5004
    • Reward points : 0
    • Joined: 2019/02/28 09:27:55
    • Status: offline
    • Ribbons : 22
    Re: RTX Titan for machine learning 2019/12/19 07:02:35 (permalink)
    With that case setup and front mount AIO you're going to be starving the GPU's for air in a major way, I'd look into something with 3 unobstructed intake fans and a top mount AIO spot.
    #4
    rhussain
    FTW Member
    • Total Posts : 1679
    • Reward points : 0
    • Joined: 2009/07/17 06:58:49
    • Location: Infinite Loop
    • Status: offline
    • Ribbons : 7
    Re: RTX Titan for machine learning 2019/12/22 18:24:44 (permalink)
    Thanks for all your suggestions folks. I'll rethink the power supply and cooling. I might consider getting one card at the moment for my workload and see how it performs. Does anyone think a single titan rtx will outperform a 1080ti or 2080ti when it comes to training neural network models? I am particularly testing with tensorflow.

     



    #5
    MasterMiner
    SSC Member
    • Total Posts : 713
    • Reward points : 0
    • Joined: 2018/01/25 21:39:33
    • Status: offline
    • Ribbons : 2
    Re: RTX Titan for machine learning 2019/12/22 18:34:00 (permalink)
    rhussain
    Thanks for all your suggestions folks. I'll rethink the power supply and cooling. I might consider getting one card at the moment for my workload and see how it performs. Does anyone think a single titan rtx will outperform a 1080ti or 2080ti when it comes to training neural network models? I am particularly testing with tensorflow.


    A single Titan RTX will blow anything any single GeForce card can do out of the water. But if $2500 is your hard dollar limit I’d go with 2 stock (Founder Edition clones) of the 2080 ti with NVLink - that should run you about same. Slightly less pooled memory (22 vs 24 gb) but more Cuda and tensor cores with a 2 card setup.

    If you’re planning a second card in the future, then get the Titan RTX. As discussed, both Titan RTX and 2080ti are hardware locked out of pcie memory pooling (unlike previous geforce 10-series which could pool memory of as many cards as you could put on a single node). Nvidia killed that off with the RTX cards. So 48 gb (2 Titan RTX) cards with nvlink is kinda your limit for model size.

    I used to mine. Now I compute.
    #6
    Jump to:
  • Back to Mobile