Nvidia Announces PCI-Express version of Tesla V100 accelerator

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Jun 21, 2017.

  1. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    41,325
    Likes Received:
    9,482
    GPU:
    AMD | NVIDIA
  2. Spets

    Spets Ancient Guru

    Messages:
    3,099
    Likes Received:
    215
    GPU:
    RTX 3090
    I can't wait for a GeForce variant :D
     
  3. Stormyandcold

    Stormyandcold Ancient Guru

    Messages:
    5,656
    Likes Received:
    343
    GPU:
    MSI GTX1070 GamingX
    Nice!

    •7 teraflops double-precision performance, 14 teraflops single-precision performance and 112 teraflops half-precision performance with NVIDIA GPU BOOST™ technology
    •16GB of CoWoS HBM2 stacked memory, delivering 900GB/sec of memory bandwidth
    •Support for PCIe Gen 3 interconnect (up to 32GB/sec bi-directional bandwidth)
    •250 watts of power

    112 Tflops of 16bit performance? Really Nvidia? That can't be real!?!
     
  4. SirDremor

    SirDremor Master Guru

    Messages:
    586
    Likes Received:
    0
    GPU:
    Nvidia GTX 1050
    Nvidia continues its mega streak after the beautifully managed Pascal lineup!

    Can't wait for Volta GeForce!
     

  5. Exascale

    Exascale Banned

    Messages:
    390
    Likes Received:
    8
    GPU:
    Gigabyte G1 1070
    There wont be one. There will be a Quadro V100 like the Quadro P100. It would play games too.
     
  6. Spets

    Spets Ancient Guru

    Messages:
    3,099
    Likes Received:
    215
    GPU:
    RTX 3090
    Tensor cores.

    Okay.. GV102 then. Think we know what I meant ;)
     
    Last edited: Jun 21, 2017
  7. Agent-A01

    Agent-A01 Ancient Guru

    Messages:
    11,404
    Likes Received:
    921
    GPU:
    ASUS 3080 Strix H20
    If it can clock to 2ghz like pascal it will be ~21 tflops :nerd:
     
  8. Truder

    Truder Maha Guru

    Messages:
    1,486
    Likes Received:
    438
    GPU:
    Sapphire Fury Nitro
    Yeah that's the ASIC "Tensor core" throughput - so that's for specific workloads to do with machine learning - it's a bit of a bullsh*ty marketing technique to make it sound way better then it really is but then again, for the target audience, that's all that's needed.

    The actual FP16 throughput is 28 TFLOPS
     
  9. Loophole35

    Loophole35 Ancient Guru

    Messages:
    9,793
    Likes Received:
    1,148
    GPU:
    EVGA 1080ti SC
    If it actually works then it's not a BS marketing ploy. Wonder how your post would have been worded had this been a Vega FE news article?
     
  10. Truder

    Truder Maha Guru

    Messages:
    1,486
    Likes Received:
    438
    GPU:
    Sapphire Fury Nitro
    Are you trying to suggest I'm being biased? Come on Loophole don't start anything like that....

    It's just the wording that isn't really accurate as such, it doesn't have 16bit 112 TFLOPS for everything, just for application specific workflows- Tensor - but that's what it's built for, machine learning.
     

  11. Loophole35

    Loophole35 Ancient Guru

    Messages:
    9,793
    Likes Received:
    1,148
    GPU:
    EVGA 1080ti SC
    Just read the article a bit closer the bullet point is wrong anyway the 112 TFlop is for the tensor core only. Still calling it marketing BS is salty AF.

    Isn't AMD trying to break into deep learning now too. Wonder if with their hopefully increased revenue from Ryzen and the fact that Vega seems to be a hit with OEM's if they will do something similar with their version of a tensor core?

    I'm more interested to see power draw and die size on the GeForce version.
     
    Last edited: Jun 21, 2017
  12. Truder

    Truder Maha Guru

    Messages:
    1,486
    Likes Received:
    438
    GPU:
    Sapphire Fury Nitro
    I just don't like any form of mis-representation, the bullet point should say that tensor core throughput is 112 TFLOPS, it looks like it's describing general compute performance.

    Kinda like Broadband ISPs advertising upto (for the sake of argument) 100mbit but the actual speed is only 10% of that because of limitations that are detailed in small print. (Unless it's the article/news source that's written wrong?).

    The big businesses investing in this technology know exactly what they're getting though so it's not really an issue, this technology isn't for a general consumer pleb like me and others around here.
     
  13. KingK76

    KingK76 Member Guru

    Messages:
    106
    Likes Received:
    11
    GPU:
    Pascal Titan X Under H2O
    Hilarious! I did the exact same thing you must have done after reading this article. I came up with the same 21Tf figure you did... The cool thing is though that at 12nm the Geforce variant (GV102) may clock even higher then 21MHz... especially considering that the Geforce catd won't have to waste power on Tensor cores...
     

Share This Page