Page 1 of 9 123 ... LastLast
Results 1 to 10 of 87

Thread: Nvidia Volta

  1. #1
    Senior Member
    Join Date
    Oct 2011
    Posts
    1,355

    Nvidia Volta

    Nvidia talking about Volta

    https://devblogs.nvidia.com/parallel.../inside-volta/

    https://videocardz.com/69378/nvidia-...120-cuda-cores

    12nm, 5376 CUDA FP32, 15.0 TF, 1455 MHz, 815 mm2 die size, new tensor cores, NV link 2.0 , 16GB HBM2, TDP 300W

  2. #2
    Senior Member
    Join Date
    Nov 2011
    Posts
    2,591
    815mm^2, I hope it comes with a giant H on top to signal that helicopters are safe to land there.
    Within 27% margin of error and 3mm off my estimate of something 4mm, good enough to predict the future

  3. #3
    Senior Member
    Join Date
    Jan 2012
    Posts
    755
    Looks like a beast for certain use cases, a lot of wasted die space for others.

  4. #4
    Banned
    Join Date
    Oct 2010
    Posts
    490
    Didn't I just ban you for acting like this?

  5. #5
    Senior Member
    Join Date
    Jan 2012
    Posts
    755
    Quote Originally Posted by raghu78 View Post
    what an ignorant statement. This is a datacenter GPU focussed at Deep learning / AI, HPC and GPU compute and only sold in supercomputers, GPU compute servers . You will see a GV102 with the same 5376 cores (probably 5120 enabled) which is stripped of all the HPC features and sold as the next Titan / Geforce 2080 Ti. Die size should be 550-600 sq mm. Launch Q1 2018.
    Ignorant statement, huh? You just said the exact same thing that I said.

  6. #6
    Senior Member
    Join Date
    Dec 2012
    Posts
    7,390
    So, like I was saying, Nvidia has the die area to burn, and are willing to burn it.

    Volta is going to win the crown. Unless Vega is something insanely special.

    I am also surprised TSMC has a reticle that is large enough to make this. Someone who knows more about process can step in?

    I recall usually TSMC pushed the reticle/max size at around 600mm^2. I recall Intel having made some "over 600mm^2" chips before (not that Intel officially shared the die size) but that's a special case.


    I would also like to point out, this chip delivers a whopping 3.2% density improvement! So 12m just being 16FF+++ is confirmed IMHO.
    -Q

  7. #7
    Senior Member
    Join Date
    Jun 2010
    Posts
    4,227
    Nothing about Desktop Volta? I suppose this means H1 2018 with GDDR6.
    P100 was 610mm^2 and P102 475mm^2, 135mm^2 saved, not sure why people think Volta 102 will have 5372 shaders with a die size of 500-600mm^2.

  8. #8
    Senior Member
    Join Date
    Jul 2011
    Posts
    2,049
    Aaaannddd... Nvidia made a bigger die. Just like before. Nothing new here (except the tensor core, gonna need to read more about that)

  9. #9
    while Raghu got banned, I think his numbers are off. When you look at GP100 vs GP102, you lose a third of the shaders(the dp ones)and go from 610mm^2 to 471mm^2. That is saving ~20% die size, dropping DP, some other stuff maybe but adding some space back in for the memory controller. A broadly similar die space saving of 20% and again going from HBM2 to a GDDR based memory controller would go from 810mm^2 to around 650mm^2.

    While a 810mm^2 core can be frankly any price and get sold to HPC companies without a problem, 650mm^2 in the consumer space is asking for a lot of problems with some huge costs. I can't see it being viable, and unless GDDR6 brings insane power savings, then it's going to struggle to compete inside a reasonable power bracket within the consumer space. It's also why I can't see Nvidia having a core anywhere near as big as 650mm^2 with gddr6, even if the memory chips presumably run a little lower voltage and significantly less power than gddr5 or even 5x, it's the IO side that HBM gains from and the more bandwidth you want the bigger the power from the memory controller is.

  10. #10
    Senior Member
    Join Date
    Dec 2012
    Posts
    7,390
    Quote Originally Posted by Drunkenmaster View Post
    while Raghu got banned, I think his numbers are off. When you look at GP100 vs GP102, you lose a third of the shaders(the dp ones)and go from 610mm^2 to 471mm^2. That is saving ~20% die size, dropping DP, some other stuff maybe but adding some space back in for the memory controller. A broadly similar die space saving of 20% and again going from HBM2 to a GDDR based memory controller would go from 810mm^2 to around 650mm^2.
    650mm^2 is about right if using GDDR5X/GDDR6.
    -Q

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •  
WordPress Appliance - Powered by TurnKey Linux