AMD Polaris 10 GPU To Offer Near 980 Ti Performance For 299 USD?

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Apr 26, 2016.

  1. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    Okay slow down a minute, I had an argument of primitive discard engines (PDE) especially regarding Hitman.

    Slow down a minute.

    Primitive discard means you can cull unseen triangles generally speaking, this has been a feature since Kepler if I'm not mistaken.

    This contributes a big part to the geometry performance on NV architectures being so much higher than on GCN.

    Now, 2560 cores at 1500mhz are 7.7tflops, which would put it quite comfortably in 980ti and fury x range

    Leaving aside the fact that we have no indication it will clock that high, it will very likely be bandwidth limited on a 256-bit bus, and nothing so far points to Polaris launching with gddr5x.

    Let's go back to hitman for a minute, IF the PDE on Polaris makes a big difference vs Hawaii for example (in terms of performance) i would be very surprised. Why, you ask ?

    Well, if PDE makes a noticeable difference this means geometry performance is a limiting factor. If geometry performance is a limiting factor there's no reason why the 970/980/980ti are performing badly relative to their AMD competition, except intentional 'gimping'. Now as you know I'm not a fan of this conclusion, and I prefer pretty much any explanation that doesn't involve a hardware conspiracy.

    One thing I've mentioned in the past, but I'm not entirely sure how many of you took note of it is that based on AotS (despite what AMD claim regarding Hitman, it's clearly the game that benefits most from Async Shaders) AMD actually is at a disadvantage in terms of shader utilization compared to maxwell.

    What does this mean really, you ask ?

    Well it means that if we have two GPUs with similar floating point throughput and we consider the use of asynchronous shaders (meaning asynchronous, (possibly) concurrent execution of graphics + compute) then ideally one of two things will happen

    1. No performance gain

    2. Performance gain.

    In the second case, this performance gain necessarily implies that the shader utilization without AS(async shader) was not maximum.

    If I have two GPUs both capable of 8tflops of FP32 and they both perform equally in a compute-bound benchmark (Fp32 is bottleneck) then I can conclude that their shader utilization is similar. We cannot say for sure they are both 100% utilization because unless we can actually quantify flop/s in the benchmark.

    In AotS you have this situation when comparing maxwell to gcn at similar FP32 throughput.

    If maxwell with no AS can perform like GCN with AS it means the shader utilization is similar, which necessarily implies that shader utilization WITHOUT AS is worse than on maxwell

    Raising clocks by 50% would be quite a feat, 16/14nm FF certainly wouldn't account this for entirely, there would have to be significant architectural changes to achieve this. It's certainly possible, but much more likely Polaris 10 will end up a ~130w part with 2560 SPs @ ~ 1100mhz with performance to match a 390x at half the power.
     
    Last edited: Apr 27, 2016
  2. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    And something else I feel compelled to point out, despite my being sure it will upset some people.

    But it's virtually guaranteed that this "Polaris 10 performs close to 980ti' claim is based on a single DX12 title in which a 390x pretty much matches a 980ti. (I wonder which game that is :p) and is probably compared to a stock/reference 980ti..

    Just for reference, a stock 980ti runs at a boost clock of~ 1100mhz. That's 6.2 tflops

    2560 x 2 x 1200mhz = 6.2 tflops
     
  3. Denial

    Denial Ancient Guru

    Messages:
    14,206
    Likes Received:
    4,118
    GPU:
    EVGA RTX 3080
    Is it based on a title or 3DMark? I thought the whole thing came from wccftech's "almost 4000" score but maybe slightly lower, or whatever nonsense they said.
     
  4. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,125
    Likes Received:
    969
    GPU:
    Inno3D RTX 3090
    I have no idea about that. All GPUs were doing some kind of occlusion culling since the ATI 8500 days at least, but I don't know what that means exactly, and why AMD put so much emphasis on it on their slides. It might mean that it really does something like what PowerVR is doing, and that could potentially mean huge gains per clock vs older GCN revisions. We have to wait and see.

    You might as well be right about Hitman. You're making only one mistake. Don't fall into the Flop trap :p

    It's not only what's there, but how you can tap it and in what conditions. GCN has always had higher TFlop performance, that doesn't mean a lot.
     

  5. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    I understand what you mean about the flop trap, but the only reason it didn't mean a lot when GCN had the advantage (Hawaii!) is because of a combination of what Dr. Rus is saying, and what you're saying; so a tendency for developers to make use of more pixel shaders/compute/post process vs geometry effects + potential driver optimizations for maximizing utilization

    There's no getting round a flop/s advantage, you can argue that what really matters is % utilization, and that's true, but generally speaking you need a spectacularly ****ty situation for a 10 tflop chip to be outperformed by a 8 tflop chip :p

    Afaik it's different from occlusion culling, here you are culling triangles whose area is smaller than 1 pixel for example, dont quote me on it though, i'm no expert and I didn't even google before posting this .

    GCN has been performing better (relative to nv competition) at high resolutions generally because the ratio of geometry:shader work is skewed towards the latter at high res, with this bottleneck lifted all that will happen is that shader throughput can go back to being a bottleneck even at low res high fps

    AMD claimed doubled geometry performance if I'm not mistaken, this would put it on par with maxwell
     
    Last edited: Apr 27, 2016
  6. Denial

    Denial Ancient Guru

    Messages:
    14,206
    Likes Received:
    4,118
    GPU:
    EVGA RTX 3080
    Occlusion is when you don't render targets obscured by others.

    This guy on reddit explains it pretty well:

     
  7. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    Yeah, primitive discard isn't doing that afaik, + you can do occlusion culling entirely by software no ? I meant that PD is culling small triangles
     
  8. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,125
    Likes Received:
    969
    GPU:
    Inno3D RTX 3090
    If you are culling non-visible triangles, aren't you in effect not rendering what is not visible?
     
  9. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    Yeah, if the area of a triangle is less than one pixel, or rather two of its vertices are on the same pixel you don't wanna perform any computation on it, it's just wasteful. It makes perfect sense in the context of tesselation, you have a model that takes up X space when you're far away (on the frame), many of the details of the model will be lost to pixel count. When the thing takes up more space those extra triangles will go towards more accurate shadowing etc etc

    Afaik this has been GCN's main hurdle with geometry

    Now I think about it, imagine AotS had better models, more detailed terrain, with those same numbers of objects onscreen, we'd be seeing very different benchmark data :p
     
    Last edited: Apr 27, 2016
  10. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,125
    Likes Received:
    969
    GPU:
    Inno3D RTX 3090
    Then you would have other stuff too. Wouldn't be compute-related things more heavy too? Raw rendering surfaces also. I don't know. The fact is that all of gaming has to be optimized for GCN, and it's ironically NVIDIA that made the first compute-oriented GPU (Fermi).
     

  11. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    Man it's not about being less compute-oriented, it's simply less floating point performance, AMD has always had the lead in this recently
     
  12. Fender178

    Fender178 Ancient Guru

    Messages:
    4,194
    Likes Received:
    213
    GPU:
    GTX 1070 | GTX 1060
    I don't see this happening because AMD had a hard time keeping up with the 980 ti. Yes the 980 ti and the Fury X traded blows in some games and in others the 980 ti came out on top. However it would be nice if AMD could accomplish this.
     
  13. Undying

    Undying Ancient Guru

    Messages:
    25,358
    Likes Received:
    12,756
    GPU:
    XFX RX6800XT 16GB
    Problem is that we dont have many dx12 games atm, if we had like 10-20 games Fury X would come on top easily.

    I was just watching comparison between 980ti and Fury X in Hitman second mission. At times Fury X has 20fps more.
     
  14. Reddoguk

    Reddoguk Ancient Guru

    Messages:
    2,661
    Likes Received:
    594
    GPU:
    RTX3090 GB GamingOC
    300$ is around 200 in the UK right now, pounds not that strong atm.

    Still if it's true pricing and i'll add in some extras for tax and shipping and i'll add a bit more for none reference, so say 300 pound.

    If i could get a gfx card for £300 that beats a 980Ti then it would be happy days for gamers.

    The 980 G1 i have was an eye watering 600 quid at release and must be still worth a few quid. I'd sell it at the drop of a hat for a 980Ti beater any day even if i'm a self confessed fan boi of Nvidia. Must be worth 200 pound at least. So i'm getting a 980Ti eater for a hundred pounds, happy days...
     
  15. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    8,125
    Likes Received:
    969
    GPU:
    Inno3D RTX 3090
    Link?
     

  16. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000

    Dx11 already has AMD very ahead, this has nothing to do with dx12, it's just about hitman
     
  17. Undying

    Undying Ancient Guru

    Messages:
    25,358
    Likes Received:
    12,756
    GPU:
    XFX RX6800XT 16GB
    I'll point out that first mission did favor AMD even more, second mission not so much but some scenes (at the end) like i said, 20+fps difference.

    https://www.youtube.com/watch?v=Do-Ex7VtzfA
     
  18. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    It's also a reference 980ti, what is he using to for an overlay in dx12

    I checked his other videos to see clocks, 1215 on the 980ti, 1180 on titan x, 1253 gtx 980
     
    Last edited: Apr 27, 2016
  19. Ryu5uzaku

    Ryu5uzaku Ancient Guru

    Messages:
    7,547
    Likes Received:
    608
    GPU:
    6800 XT
    The "small" die that Fury X is on won't oc is the reason why it is so behind 980 ti at times. If the die size was more or less similar in density to 980 ti the situation would be tad different. Having higher density does no favors really.
     
  20. anticupidon

    anticupidon Ancient Guru

    Messages:
    7,878
    Likes Received:
    4,127
    GPU:
    Polaris/Vega/Navi
    Sorry to be late at this party and sorry for asking
    Will Polaris 10 have HBM2 or this will be on Vega only ?
     

Share This Page