NVIDIA Volta Rumor Thread

Discussion in 'Nvidia' started by Arachnotronic, Feb 11, 2017.

  1. Samwell

    Samwell Member

    Joined:
    May 10, 2015
    Messages:
    192
    Likes Received:
    42
    Volta is a pure compute oriented architecture with no new graphic features and won't come to gaming. Ampere will now be Volta - HPC Stuff + more advanced Graphic Features. Since the differences between HPC and gaming are growing, Nvidia decided to call the architecture different.
     
  2. PeterScott

    PeterScott Senior member

    Joined:
    Jul 7, 2017
    Messages:
    974
    Likes Received:
    525
    It's really just a name change, which is kind of irrelevant. GTC is at the end of March, and rumors indicate they won't be shipping for months after that. Well over 2 year on the same series.

    NVidia really gets to sit back and milk GPU designs these days.
     
  3. Ajay

    Ajay Diamond Member

    Joined:
    Jan 8, 2001
    Messages:
    4,376
    Likes Received:
    1,151
    Yes. It does look like NV will be forking it's graphics GPUs to 'compute' and 'gaming gfx'. Makes sense given how compute has developed into a financially significant market. I imagine there will be a fair bit of overlap in R&D to keep costs down - it'll be interesting to see if the SM structures are pretty much the same.
     
  4. Bouowmx

    Bouowmx Senior member

    Joined:
    Nov 13, 2016
    Messages:
    674
    Likes Received:
    274
    Gaming and HPC architecture has already been forked: GP100 vs GP<everything else>. I suppose this rumor is about the formal recognition of the fork.
     
  5. Bier667

    Bier667 Member

    Joined:
    Oct 31, 2017
    Messages:
    32
    Likes Received:
    1
    WTF?

    From the german article quoted on newssites:
    https://www.heise.de/ct/ausgabe/2017-16-Von-Mode-und-neuen-Kleidern-3774615.html


    Volta -> Ampere

    And as Bouowmx pointed out allready Nvidia has the Tesla only G100 (1:2 DP,Tensorcores...) and G102/104 HPC parts for Quadro, Tesla and Gaming (1:32DP , no Tensorcores...).
     
  6. Trumpstyle

    Trumpstyle Member

    Joined:
    Jul 18, 2015
    Messages:
    41
    Likes Received:
    12
    Most likely "Ampere" is volta replacement which will be announced on GTC 2018. It will be 10 nm and replace GV100.

    I still believe we will see volta geforce Gpus in feb-march with gddr6. Keep in mind regarding this rumor that Nvidia never launches geforce cards at GTC events. So this Ampere rumor is probably about a replacement card for GV100 on 10 nm.
     
  7. Bier667

    Bier667 Member

    Joined:
    Oct 31, 2017
    Messages:
    32
    Likes Received:
    1
    GV102 for Tesla/Quadro/Titan could be a possibility. As for the Geforce Lineup: i have my deepest doubts.
     
  8. Trumpstyle

    Trumpstyle Member

    Joined:
    Jul 18, 2015
    Messages:
    41
    Likes Received:
    12
    After further thoughts, Ampere is probably just the successor to Volta. It will be announced at GTC 2018 but we won't see any gpus from this architecture intill 2019 and it will be on 7nm. Volta Geforce still will be released at feb-march with gddr6.
     
  9. imported_bman

    imported_bman Senior member

    Joined:
    Jul 29, 2007
    Messages:
    257
    Likes Received:
    53
    Ampere just sounds like the consumer fork of Nvidia's post-Pascal architecture, where as Volta is the HPC variant with tensor cores. Also the names Volta and Ampere are perfect for each fork of the post-Pascal architecture given the relationship between voltage and current.
     
    William Gaatjes likes this.
  10. Trumpstyle

    Trumpstyle Member

    Joined:
    Jul 18, 2015
    Messages:
    41
    Likes Received:
    12
    Another source saying Volta is being dropped for Ampere.

    http://www.3dcenter.org/news/nvidia...bringt-im-q22018-gleich-die-ampere-generation

    Used a translation tool so everything I write might not be 100% accurate. A natural born speaking german can probably do a better job translating than me.

    But the article states that they seen the internal roadmap and volta has been dropped and replaced by Ampere. Ampere will be using TSMC 10nm.

    Volta gaming gpus was planned but 12nm from TSMC was just not good enough for a decent performance gain they think. Most likely (my opinion) Nvidia saw they got get better margins using 10nm over 12nm. As a GV104 with 400 mm^2 die size would probably give a 50-70% performance increase. But these ampere gpus should be released in q2 2018.

    Anyway with 2 sites saying Volta is being dropped for ampere it most likely true. 1 site saying q2 release and another says expect announcement at gtc 2018 (march 26).

    Only thing that is unclear if Ampere will replace GV100 (article does not say), it would make sense go to 10nm for better yields and margins.
     
    Jaskalas and Bouowmx like this.
  11. Samwell

    Samwell Member

    Joined:
    May 10, 2015
    Messages:
    192
    Likes Received:
    42
    @Trumpstyle

    No, they haven't seen any roadmaps. They're just guessing and expecting a node jump, but have no information about it. The internal roadmap stuff ist about the name Ampere and other people seeing this roadmaps. But at the moment only the name is clear.

    Beside the name the only info which we know is the stuff from Drive PX Pegasus. Pegasus seems to have an Ampere GPU which gives us the information, that some Ampere gpu is also having tensor cores. These Tensor cores will probably work with lower precision as the volta tensor cores, so they will be smaller. V100 is FP16 Multipli and FP32 accumulate, which can be used for training. Ampere might have FP8/FP16 tensor cores for inference. You can even calculate a possible shader number by looking at Pegasus TOPS ;)

    I don't believe in 10nm. Everything indicates 12nm at the moment. Why would TSMC make some changes for Nvidia for 12FFN if it's only for V100 and Xavier (Even if the process changes are very small)? I believe the whole next gen will be 12nm this way Nvidia convinced TSMC to make a costum node for Nvidia.
     
    #336 Samwell, Nov 20, 2017
    Last edited: Nov 20, 2017
  12. Bier667

    Bier667 Member

    Joined:
    Oct 31, 2017
    Messages:
    32
    Likes Received:
    1
    Ampere counters Intel exascale.


    I don't know why peeps here dun get the memo.
     
  13. Bier667

    Bier667 Member

    Joined:
    Oct 31, 2017
    Messages:
    32
    Likes Received:
    1
    Your comment is so full of (redacted) as with the president of the undited states. We from europe have a different taste of what could happend or not: i think the Balkan makes us to this.


    Anyway: you american guys are rly funny.

    /edit
    there ain't a 10nm node.

    Moderator Edit:
    Profanity and personal attacks are not allowed in technical forums. Political discussion belongs in the Politics & News forum. Please keep this in mind when posting outside of the Social forums. -IEC
     
    #338 Bier667, Nov 29, 2017
    Last edited by a moderator: Nov 29, 2017
  14. Bouowmx

    Bouowmx Senior member

    Joined:
    Nov 13, 2016
    Messages:
    674
    Likes Received:
    274
    #339 Bouowmx, Dec 7, 2017
    Last edited: Dec 7, 2017
  15. Arachnotronic

    Joined:
    Mar 10, 2006
    Messages:
    11,475
    Likes Received:
    1,805
  16. imported_bman

    imported_bman Senior member

    Joined:
    Jul 29, 2007
    Messages:
    257
    Likes Received:
    53
    https://www.nvidia.com/en-us/titan/titan-v/?ncid=so-you-tnv-29059#shop

    3 Grand! Though I am sure there is someone on these forums that will buy one.



    Specs:
    Graphics Processing Clusters 6
    Streaming Multiprocessors 80
    Cuda Cores 5120
    Texture Units 320
    Base Clock 1200 MHz
    Boost Clock 1455 MHz
    Memory Clock 850 MHz
    Memory Data Rate 1.7 Gbps
    L2 Cache Size 4608K
    Total Memory 12288 MB HBM2
    Memory Interface 3072-bit
    Total Memory Bandwidth 652.8GB/s
    Texture Rate (Bilinear) 384 GigaTextels/sec
    Fabrication Process 12nm
    Transistor Count 21.1 Billion
    Connector 3 x Displayport, 1 x HDMI
    Power Connectors One 6-pin, One 8-pin
    Recommended Power Supply 600 Watts
    Thermal Design Power (TDP) 250 Watts
    Thermal Threshold 91C
     
    #341 imported_bman, Dec 7, 2017
    Last edited: Dec 7, 2017
  17. Bouowmx

    Bouowmx Senior member

    Joined:
    Nov 13, 2016
    Messages:
    674
    Likes Received:
    274
  18. Bouowmx

    Bouowmx Senior member

    Joined:
    Nov 13, 2016
    Messages:
    674
    Likes Received:
    274
  19. slashy16

    slashy16 Member

    Joined:
    Mar 24, 2017
    Messages:
    100
    Likes Received:
    39
    Its sad how far ahead Nvidia is now. Hopefully AMD catches up and doesn't fall 3 generations behind.
     
  20. Bondrewd

    Bondrewd Senior member

    Joined:
    Oct 16, 2017
    Messages:
    360
    Likes Received:
    114
    Hurray, non-toy Titan is back!
    For 3grand anyway.
    That's not gonna happen though.
     
  21. tamz_msc

    tamz_msc Golden Member

    Joined:
    Jan 5, 2017
    Messages:
    1,630
    Likes Received:
    1,488
    AMD has the same chances with a high-end GPU as Microsoft has in selling the Xbox in Japan.

    On point though, it seems that NVIDIA has finally stopped being wishy-washy about what the Titan range is supposed to be. This is exactly the same as the Tesla V100 except for the cut down memory.
     
  22. Carfax83

    Carfax83 Diamond Member

    Joined:
    Nov 1, 2010
    Messages:
    5,533
    Likes Received:
    330
  23. moonbogg

    moonbogg Diamond Member

    Joined:
    Jan 8, 2011
    Messages:
    9,434
    Likes Received:
    886
    I guessed $2,000 if Volta came this year. I thought I had reached high enough, but I guess not. $3,000, wow. It makes sense though considering how much faster it is and how soon it came. Its an early next gen Titan and doesn't seem to be using a small die. That's a big die real deal card right there. Hopefully it kicks major ass and gives us a clue about the next Ti's performance, so long as that Ti doesn't cost $1,500 or something (probably will).

    I expect at least one particular forum member to buy two of them with that $600 SLI bridge. Let us know when you got those benchies! You know who you are and so do we.
     
  24. Bondrewd

    Bondrewd Senior member

    Joined:
    Oct 16, 2017
    Messages:
    360
    Likes Received:
    114
    Its barely faster than GP102 in FP32.
    For all intends and purposes, that's not a consumer card.
     
  25. moonbogg

    moonbogg Diamond Member

    Joined:
    Jan 8, 2011
    Messages:
    9,434
    Likes Received:
    886
    So it won't game a lot faster than a 1080ti?