Page 27 of 27 FirstFirst ... 17252627
Results 261 to 269 of 269

Thread: Nvidia Volta

  1. #261
    Senior Member
    Join Date
    Jun 2010
    Posts
    1,360
    Quote Originally Posted by Bluefoot View Post
    What are you talking about? Are you seriously trying to say that either Fiji was 14nm or Polaris / Vega are only comparable to Maxwell? If so, there's no point talking. Fiji is faster than Maxwell, and Maxwell is a long, long way behind performance of AMD 14nm products. Its efficiency is about the same as Pascal (maybe slightly lower) because the latter is essentially the former at much higher clocks, outside of P100.
    I am comparing Polaris to Maxwell, where Polaris offers only slightly better efficiency and performance than Maxwell despite 2 node advantage (14 vs 28). Perhaps AMD is maybe half a node ahead here, not more - there is a much larger gap between Pascal and Polaris. At stock settings of course. Sure, you could claim AMD could have set 480 at lower frequency and have better perf/W... but they didn't.
    This about half a node seems the same as performance difference between 16nm and 12nm from TSMC. This is leading me to guess 7nm Navi will be comparable to 12nm Volta in gaming performance, assuming similar efficiency gains of Navi and Volta (vs Vega and Pascal).

    @LurkingLeopard
    Well, what would NV gain by that? They are de-facto GPU option already and can sell same sized GPU for twice as much as AMD (excluding mining prices). They can't kill AMD or even cut them off completely from gaming either, as AMD retains consoles and should be doing fine with APUs.

    I maintain my guess that even consumer Volta will be slightly AI compute oriented, at the expense of highest possible gaming performance - because it is going to be fast enough anyway.

  2. #262
    Senior Member
    Join Date
    Jan 2010
    Posts
    3,571
    14 isn't 2 nodes ahead of 28 except in MarketingWorld.

    Everything indicates that Polaris was a cost-down design - small dies for competitive prices, rather than a performance design. They left Fiji to compete at the high end, but Nvidia likely beat AMD's expectations, and that screwed up some of the Polaris aims.

    It remains to be seen what a 12nm Vega will bring for AMD. +10% clocks isn't going to change anything without a fix to improve gaming performance to better meet theoretical TFLOPS (the lack of scaling from V56 to V64 when everything else is the same is a big flag to there being an issue, but it also means that Vega 11 may not be affected and thus outperform on a perf/CU basis).

    What GV104 will bring beyond GP104 is unknown. Trained AI is going to be big in the future in gaming, and it will only take off when integrated into a GPU without affecting graphical performance. In addition I can see a market for Trainable AIs, for top gamers to train their playstyles, for others to play against - this seems like a good thing for a GV102 product to provide, but more likely will be a generation or two beyond Volta.

  3. #263
    Senior Member
    Join Date
    Dec 2010
    Location
    Switzerland Geneva Cern
    Posts
    2,933
    Quote Originally Posted by JeeBee View Post
    14 isn't 2 nodes ahead of 28 except in MarketingWorld.

    Everything indicates that Polaris was a cost-down design - small dies for competitive prices, rather than a performance design. They left Fiji to compete at the high end, but Nvidia likely beat AMD's expectations, and that screwed up some of the Polaris aims.

    It remains to be seen what a 12nm Vega will bring for AMD. +10% clocks isn't going to change anything without a fix to improve gaming performance to better meet theoretical TFLOPS (the lack of scaling from V56 to V64 when everything else is the same is a big flag to there being an issue, but it also means that Vega 11 may not be affected and thus outperform on a perf/CU basis).

    What GV104 will bring beyond GP104 is unknown. Trained AI is going to be big in the future in gaming, and it will only take off when integrated into a GPU without affecting graphical performance. In addition I can see a market for Trainable AIs, for top gamers to train their playstyles, for others to play against - this seems like a good thing for a GV102 product to provide, but more likely will be a generation or two beyond Volta.
    After seeing that the vega 56 has the same perf / clock with vega 64 its obvious that amd still has scalling and utilization issues with vega. Vega 64 should be 14% faster than vega 56 on the same clocks! That would put vega 64 comfortably ahead of 1080 even with current non optimal drivers...

    Also according to techpowerup database liquid cooled vega would be as fast as titan pascal with a 14% boost.
    Current rig: i5 3570 @3.8GHz, Asus 7750, Dell 24'' 2560X1440, Corsair Vengence 16GB Ram DDR3 @1600MHz, Chieftec 550W PSU, 480GB OCZ SSD + 500GB Seagate + 1TB WD hd, Windows 10 64-bit

  4. #264
    Senior Member
    Join Date
    Jun 2010
    Posts
    1,360
    Quote Originally Posted by JeeBee View Post
    14 isn't 2 nodes ahead of 28 except in MarketingWorld.

    Everything indicates that Polaris was a cost-down design - small dies for competitive prices, rather than a performance design. They left Fiji to compete at the high end, but Nvidia likely beat AMD's expectations, and that screwed up some of the Polaris aims.

    It remains to be seen what a 12nm Vega will bring for AMD. +10% clocks isn't going to change anything without a fix to improve gaming performance to better meet theoretical TFLOPS (the lack of scaling from V56 to V64 when everything else is the same is a big flag to there being an issue, but it also means that Vega 11 may not be affected and thus outperform on a perf/CU basis).

    What GV104 will bring beyond GP104 is unknown. Trained AI is going to be big in the future in gaming, and it will only take off when integrated into a GPU without affecting graphical performance. In addition I can see a market for Trainable AIs, for top gamers to train their playstyles, for others to play against - this seems like a good thing for a GV102 product to provide, but more likely will be a generation or two beyond Volta.
    7 isn't really 2 full nodes ahead of 14 either These names are a mess, but after all the marketing, jump from 28 -> 14 is similar to 14 -> 7 in most of stuff (performance, power, density). More than a single node jump, less than 2.

    With 12nm I was thinking Volta, isn't that made on TSMC's 12nm, and should be competing with 7nm Navi? (assuming Navi launches when it is supposed to, based on old AMD roadmaps. With delays like on Vega, Volta could be without competition and Navi to be battling what comes after Volta)

    AI in games will happen, but I don't think it will happen before X2 and PS5 get released with AI chips inside. Definitely a few generations after Navi and Volta. I am considering other AI stuff now, as well as offering hardware to make such AI games happen.

  5. #265
    Senior Member
    Join Date
    Sep 2009
    Posts
    2,064
    Quote Originally Posted by Nerdmaster View Post
    Vega 64 should be 14% faster than vega 56 on the same clocks! That would put vega 64 comfortably ahead of 1080 even with current non optimal drivers...

    Also according to techpowerup database liquid cooled vega would be as fast as titan pascal with a 14% increase in shader count.
    Theoretical performance numbers are meaningless as the real world doesn't work this way. Many gaming benchmarks are at least somewhat CPU limited. A good GPU would be fortunate to get more than a 10% performance increase with a 14% increase in core count.

    The footnotes on that TechPowerUp GPU database state: "Based on TPU review data: "Performance Summary" at 1920x1080". A GTX 1080 has 33.3% more cores than a GTX 1070 (2560 vs 1920). Yet in the TechPowerUp Vega 64 review the GTX 1080 is only 15.6% faster (104%/90%) than the GTX 1070 @ 1920x1080.

    So a "14% boost" is really going to require something like 30% more clock speed or cores (assuming same scaling as NVIDIA cores) to hit that 14% higher overall benchmark performance.

    Vega has a long way to go just to catch GP102.
    Last edited by eRacer; 21st September 2017 at 07:51 AM.

    Primary System: Core i7-6700K - GTX 1060 6GB
    Secondary System: AMD FX-6300 - Radeon RX 470
    Gaming-Only System: Core i5-4670K - Radeon RX 480 8GB

  6. #266
    Quote Originally Posted by Zizy View Post
    I Well, what would NV gain by that? They are de-facto GPU option already and can sell same sized GPU for twice as much as AMD (excluding mining prices). They can't kill AMD or even cut them off completely from gaming either, as AMD retains consoles and should be doing fine with APUs.
    Nvidia and AMD both still make most money with gaming. As long as AMD can only compete in that segment with horribly overpriced silicon sold at low prices, they're starving the beast.

    I maintain my guess that even consumer Volta will be slightly AI compute oriented, at the expense of highest possible gaming performance - because it is going to be fast enough anyway.
    I also think that there'll be AI related stuff in gaming GPUs. But not at the cost of gaming performance.

    Gamers need a good reason to upgrade.

  7. #267
    Senior Member
    Join Date
    Dec 2012
    Posts
    7,451
    14nm is 20nm+/++ with FinFET
    -Q

  8. #268
    Senior Member
    Join Date
    Dec 2010
    Location
    Switzerland Geneva Cern
    Posts
    2,933
    Quote Originally Posted by eRacer View Post
    Theoretical performance numbers are meaningless as the real world doesn't work this way. Many gaming benchmarks are at least somewhat CPU limited. A good GPU would be fortunate to get more than a 10% performance increase with a 14% increase in core count.

    The footnotes on that TechPowerUp GPU database state: "Based on TPU review data: "Performance Summary" at 1920x1080". A GTX 1080 has 33.3% more cores than a GTX 1070 (2560 vs 1920). Yet in the TechPowerUp Vega 64 review the GTX 1080 is only 15.6% faster (104%/90%) than the GTX 1070 @ 1920x1080.

    So a "14% boost" is really going to require something like 30% more clock speed or cores (assuming same scaling as NVIDIA cores) to hit that 14% higher overall benchmark performance.

    Vega has a long way to go just to catch GP102.
    First of all you are talking about cpu bottleneck which is completely irrelevant to the scalling issues of vega architecture.
    1080 has 21% extra performance in 2560x1440 as well as 22% extra performance on 4k resolution.

    A 33% of extra cores and a minor clock speed increase resulting in ~20% extra performance.

    Vega 64 has 14% more CUs than vega 56 but 0 performance improvement when compared to vega 56 at the same clocks when tested in multiple games and multiple resolutions in many of which a gpu bottleneck is apparently ensured.

    If the issue was cpu bottleneck the extra clocks of vega 64 wouldnt improve performance at all.
    Current rig: i5 3570 @3.8GHz, Asus 7750, Dell 24'' 2560X1440, Corsair Vengence 16GB Ram DDR3 @1600MHz, Chieftec 550W PSU, 480GB OCZ SSD + 500GB Seagate + 1TB WD hd, Windows 10 64-bit

  9. #269
    Senior Member
    Join Date
    Sep 2009
    Posts
    2,064
    Quote Originally Posted by Nerdmaster View Post
    First of all you are talking about cpu bottleneck which is completely irrelevant to the scalling issues of vega architecture.
    Yes, I am talking about the CPU bottleneck because you completely ignored it when you stated: "Vega 64 should be 14% faster than vega 56 on the same clocks!...Also according to techpowerup database liquid cooled vega would be as fast as titan pascal with a 14% boost."

    1080 has 21% extra performance in 2560x1440 as well as 22% extra performance on 4k resolution.

    A 33% of extra cores and a minor clock speed increase resulting in ~20% extra performance.

    Vega 64 has 14% more CUs than vega 56 but 0 performance improvement when compared to vega 56 at the same clocks when tested in multiple games and multiple resolutions in many of which a gpu bottleneck is apparently ensured.
    Titan Pascal and GTX 1080 Ti are also more CPU limited at 1080P than at higher resolutions. In the Vega 64 review the GTX 1080 Ti lead over Vega 64 increases from 25% at 1080P to 32% at 4K.

    Like I said in my previous post, and as evidenced by 1080/1070 at 4K, a good GPU would be fortunate to get more than a 10% performance increase with a 14% increase in core count. That would still put liquid-cooled Vega 64 overall performance below that of Titan Pascal, which is slower than GTX 1080 Ti FE, which is slower than the aftermarket GTX 1080 Ti cards that PC gaming enthusiasts have been buying for months now. The bar should be raised even higher next year when Volta arrives.

    With reference Vega 64 performance already capped by the power limit in many cases (and/or temperature in air cooled cards), I'm not sure how much more performance can be gained at the same power limit if the cores could be utilized more.
    Last edited by eRacer; 22nd September 2017 at 08:11 AM.

    Primary System: Core i7-6700K - GTX 1060 6GB
    Secondary System: AMD FX-6300 - Radeon RX 470
    Gaming-Only System: Core i5-4670K - Radeon RX 480 8GB

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