Nvidia LiveStream Event Later today

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, May 6, 2016.

  1. Ieldra

    Ieldra Banned

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

    What do you mean nothing is lost? Architecture plays a big role in clocking, the metric is meaningless.

    Most pascal cards will clock in the 1800mhz range, how exactly is a per clock metric useful to me when I compare pascal to gcn 1.2 for example.


    By this metric Hawaii is significantly better than gm200 (which I'm sure will be your endgame in the discussion) it has fewer transistors and clocks lower.

    The problem is that maxwell is designed to run at high clocks, Hawaii is not and cannot

    Until you can justify to me why an architectural performance metric would not take clocks into account I will not accept it. Are you denying that maxwell can clock inherently here, independent of cooling, power etc?

    Where does this ability to clock higher come from, and how can justify a performance metric that would not account for 50% increased performance vs hawaii?
     
  2. hexaae

    hexaae Member

    Messages:
    28
    Likes Received:
    0
    GPU:
    nVidia GTX980M 4GB G-Sync
    What sucks is:
    1. Price
    2. 256-bit Memory Interface Width
    3. still no Async Compute and real DX12 support
    Just brute force for a more efficient 980 rebrand... meh
     
    Last edited: May 7, 2016
  3. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,477
    Likes Received:
    3,181
    GPU:
    5700XT+AW@240Hz
    Seems like there are things, you'll never wrap your head around.
    My last effort is your homework: Compare Tahiti 4.3B transistors (R9-280x) vs Tonga 5.0B transistors (R9-380x), both at pretty close clocks.
    http://www.techspot.com/review/1093-amd-radeon-380x/page2.html
    Was that performance boost worth 16% more transistors? How big is average improvement between those architectures? Did AMD invest those transistors well?

    Don't be stupid. There is no endgame. Same way as I had no endgame in our benchmarking thread. (Just to shed some light into something people only speculate about.)
    I am not competing with you. I have no need for that. No reason to ruin your day or make you feel bad.
     
  4. Noisiv

    Noisiv Ancient Guru

    Messages:
    7,579
    Likes Received:
    1,014
    GPU:
    2070 Super

    not that I wish to rain on your parade :nerd:

    but do you know the guy that has been arguing with me about NV/AMD clocks?

    the fella who went on about "High clock field" being leveled on 14/16.
    said that NV 16FF+ clocks might be just a small step up from maxwell
    "All AMD needs to do is hit similar clocks like NV"

    and that that Nvidia 2GHz is out of this world. :D
     

  5. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    You know, while we frequently disagree, I took you seriously at the beginning. Lately you've been demonstrating very shallow thought processes, and an inclination towards being very thin-skinned in discussions you shouldn't be thin-skinned in.

    Now you're comparing tahiti to tonga, talking to me about how it's not a good investment of transistors...

    As if this is contradicting something I've said, it's like you just obfuscate the discussion whenever you dig yourself into a hole. Invent as many performance indices as you wish, build a GPU database in which they're ranked solely by this metric I will affectionately name Fox's Strategically Biased Performance Metric For The Betterment Of AMD's Image.

    FSBPMFTBOAI

    Nobody's heard of it, but ,with a name as convoluted as the concept behind it, it reflects on both the concept and it's author - it shall prevail, against all odds, and common sense!

    I have no doubt you will reply with some cryptic message with which you intend to lead me to assume you know something I don't, but frankly Fox your recent arguments have been as weak and unfounded as AMD's claims regarding Fiji's OC potential.

    :pc1:

    You and your ALU throughput/ ( #xtors * clocks)

    You're just rewarding GPUs with high ALU: xtor ratio with low clocks; aka GCN. D for achievement, E for effort.

    Drop the act and just outright admit you're, for some reason, enamored with wide, slow designs
     
    Last edited: May 7, 2016
  6. Endymion

    Endymion Member Guru

    Messages:
    102
    Likes Received:
    13
    GPU:
    3090 Gaming X Trio
    What about 1080 Ti?
     
  7. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,477
    Likes Received:
    3,181
    GPU:
    5700XT+AW@240Hz
    http://forums.guru3d.com/showpost.php?p=5261919&postcount=7

    Now, What do you know about 14nm GCN clocks? I mean KNOW.
    You know how GTX 1080 clocks. That's all you know about upcoming clocks. There is no difference between your and mine knowledge in GCN 1.3 clocks on 14nm. Nether of us know a thing.

    I did at least read materials from forges, to know where it can land. And I know that AMD stated that GCN 1.3 will clock reasonably higher than Tonga/Fiji.

    And somewhere there is post of mine stating that I would be surprised if desktop Pascals clocked lower than ~1500MHz. (It was in correlation to P100 official clocks.)
     
  8. HonderdProcent

    HonderdProcent Member

    Messages:
    14
    Likes Received:
    0
    GPU:
    Fury X 4GB
    Oeh this sounds interesting. I have an Fury X now with a 5K dell screen (yes, screen is for workstation use also). Most games I have to play in 2560x1440, some games I can do 5k with low/medium settings. 1070 in SLI sounds like an nice price/performance point for this. Let's see what the 4k benchmarks are going to say.
     
  9. Extraordinary

    Extraordinary Ancient Guru

    Messages:
    19,503
    Likes Received:
    1,551
    GPU:
    ROG Strix 1080 OC
  10. Koniakki

    Koniakki Ancient Guru

    Messages:
    2,843
    Likes Received:
    443
    GPU:
    ZOTAC GTX 1080Ti FE
    As much as I love reading all your comments guys, I can't stop but think that: :p

    [​IMG]
     
    Last edited: May 7, 2016

  11. Noisiv

    Noisiv Ancient Guru

    Messages:
    7,579
    Likes Received:
    1,014
    GPU:
    2070 Super

    Hey no worries, just a little stab is all :D

    I have no idea about 14LPE/LPP. I doubt it will come close to 2.1GHz in Polaris implementation. We'll see...
     
  12. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,477
    Likes Received:
    3,181
    GPU:
    5700XT+AW@240Hz
    That's your conclusion, I did not push you in anyway at all. Mine is quite different and much more complex.
    While Tonga is no improvement in older and non-complex games, it brings reasonable improvement in some of those heavy games like Witcher 3.

    Was it Worth those transistors? (You should have asked yourself: "Over what alternative?")
    Alternative was to use Tahiti again, this time on improved 28nm lithography which allowed Tonga in just 4% higher die size to get 16% more transistors and keep clock. And get 5~8% higher performance from higher clock all across board.
    So, 5~8% everywhere or up to 25% somewhere?

    That stupid metric you do not like that much would say that 'updated' Tahiti did not change at all, which would be true.
    And same way it can tell you how much Tahiti to Tonga changed.

    You probably agree with 'per transistor' part. So, why do you don't like 'per clock' part?

    How far away is it from IPC meaning for CPUs? Should IPC be ignored, because some chips are made to tick higher?
    Hell, no! It is as important as ever to compare changes with each generation. Or manufacturer.
     
  13. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    It's totally different conceptually from IPC on CPUs...

    I don't agree on the per transistor part either, we're talking about shader array performance.

    Shader array performance is a function of clock and shader count, if you divide by clock you are looking at shader:transistor ratio

    shader:transistor ratio tells you how many shaders you can expect given die area and density, and cyclic permutations of the three.

    So a chip based on uarch X with 8096 ALUs running at 1 ghz and 8bn xtors will have assigned a higher performance value by FSBPMFTBOAI than a chip based on uarch Y with 4096 ALUs at 2 ghz and 6bn xtors. ridiculous.

    FSBPMFTBOAI sucks.

    and Tonga increase geometry performance at the expense of chip complexity, which ties in nicely with my saying that AMD needs a ****ing miracle with polaris

    they need to make up for die area efficiency, and also need to decrease their ALU:transistor ratio to make up for other deficiencies.
     
    Last edited: May 7, 2016
  14. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,477
    Likes Received:
    3,181
    GPU:
    5700XT+AW@240Hz
    ^So what's your practical metric for comparison?
     
  15. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    I don't presume to have one, if it's shader array performance we're trying to quantify the closest thing to what you were saying would be to consider it as a function of xtor density. That way you can scale the value up or down to get die-specific approximations, but would still be complicated by factors like memory configuration; HBM imc is smaller and less complex than traditional gddr for example and will skew results.

    I don't claim to have a simple performance metric that works all the time, but measuring shader:transistor ratio and calling it an index of architecture performance is ludicrous

    FX: 1150mhz - 9420gflops 14.88 mn xtors per mm^2

    Titan X - 1450 mhz 8900 gflops 13.31 density

    8908/(1/13.31) vs 9420/(1/14.88)

    118,469 vs 140,199

    not to mention that actual gpu performance is more than just ALU

    edit:

    not transistor density, it's inverse.
     
    Last edited: May 7, 2016

  16. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,477
    Likes Received:
    3,181
    GPU:
    5700XT+AW@240Hz
    So, how do you say, that Pascal was step forward over Maxwell?

    What if nVidia just cloned Maxwell, onto 16nm, and higher clock is simply result of superior manufacturing process (as that's what forge claims anyway).

    What is it you are interested in knowing? Is it complex? Is it simple? Is it even remotely relevant to me wanting to know about actual changes in architecture?
    And that little thing you markes with such atrocious name/acronym definitely works for that purpose, while you have no alternative. Keep bashing...
     
  17. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    It's entirely possible that higher clock is due to the process and that's it, because Maxwell itself was tuned and optimized to maximize clock, GCN never did that.

    And here is the cryptic reply I prophesied would come earlier, this is a meaningless question, intentionally vague to absolve you of any intellectual responsibility for this terrible idea
     
  18. Fender178

    Fender178 Ancient Guru

    Messages:
    4,184
    Likes Received:
    207
    GPU:
    GTX 1070 | GTX 1060
    Man after looking at the 1070 specs I may just get that card because of the price of the card. I spent over $400 bucks for my r9 290 and maybe spending $450 for the founders edition is not that far fetched. Also Nvidia has learned their lesson over the 970. The 1070 seems like the perfect card to try DSR at either 1440p or 4k. I think $600 is way too much for me to spend right now. Either the 1080 or the 1070 is going to be a good upgrade over my r9 290 especially for Rise of the Tomb Raider @ 1440p via VSR/DSR and many more new games to come like Doom.
     
  19. Fox2232

    Fox2232 Ancient Guru

    Messages:
    11,477
    Likes Received:
    3,181
    GPU:
    5700XT+AW@240Hz
    Cryptic? Me wanting to know architecture performance, per clock and transistor? Is that cryptic, or obscuring truth?
    Architecture is architecture.

    You propose GFLOS(performance) divided by transistor density.
    On GFLOPs we agree.

    But what's actually transistor density?
    1st, it is not part of architecture. Having higher or lower transistor density does not affect performance while GPU runs same clock. It is manufacturing design choice.
    It does affect maximum clock, but in inverse function (lower the density, higher possible clock).
    It includes area of chip too, which is again not part of architecture, but manufacturing design choice (cost).

    So, your metric is particularly:
    Performance / (transistor density) =
    Performance / (transistor count / chip size) =
    - you know that division is same as multiplication by inverse value, right?
    Performance * Chip size * (1 / transistor count) =
    (Performance * Chip size) / transistor count =
    - Now, what is this construct?
    - > it divides performance by transistor count, which you disagreed with
    - > And dividend not only contains Performance, but Chip size too (which comes from transistor density => affecting clocks).
    - So, is final function some kind of power function of clock and performance divided by transistor count?
    (Performance * factor affecting clock) / transistor count =

    Does that actually look like something what can compare changes between architectures?
    And needless to say, that it makes larger chip (more costly) to look better, right?
     
    Last edited: May 7, 2016
  20. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    inverse of density, and no i'm not suggesting this is a good idea. I'm just proposing this off the top of my head as a much better idea than yours which would put two chips on the same level no matter what their clocks are

    in practice the best metric is perf/watt
     

Share This Page