Compute Performance

Compute tasks, in spite of the name, are not always purely GPU bound. Depending on the task memory bandwidth can also play a significant part – which is why memory bandwidth was the single biggest increase on the 7970 over the 6970 – and our compute benchmark end up reflecting this.

Among our tests only the DX11 Compute Shader Fluid sample fully benefits from the increased clockspeed of XFX’s factory overclock, gaining 9% over the reference 7970. Our Civilization V and SmallLuxGPU benchmarks meanwhile only gain 4-5%, and our AES benchmark only gains 2%, the latter likely due to the fact that the setup time for the program’s dataset does not decrease, only the execution time does.

Game Performance: Portal 2, Battlefield 3, Starcraft II, Civilization V Overclocking
Comments Locked

93 Comments

View All Comments

  • chizow - Tuesday, January 10, 2012 - link

    Also to this:

    "Gtx 460 and 560 ti were failure compared to AMD's offerings - in terms of performance."

    Contradict yourself much? If this is true, then you're admitting every single GPU AMD has created since G80 is a failure compared to Nvidia's offerings and directly contradicting the points you're trying to make with regard to price and performance.

    Sorry, you can't argue out of both sides of your mouth, the message just comes out a big jumbled mess.
  • wifiwolf - Tuesday, January 10, 2012 - link

    I think with this post you're just saying he was right all along.
    And I agree. It's just business, you just choose where you you want to place your product: On profits or brand.
  • chizow - Tuesday, January 10, 2012 - link

    No, I don't agree with any of that because its clearly off-base and out of place in a discussion about high-end performance parts.

    What drives pricing in this segment? This is a simple answer.

    Performance. That's all that matters.

    Performance/watt and performance/dollar are just tertiary considerations that take a back seat to secondary considerations like feature sets and application support. You win these "value" market segments not because you want to, but because you have to when you can't win without compromises.
  • wumpus - Tuesday, January 10, 2012 - link

    DPFlops/Watt matters if Fermi and Kepler were designed for GPU computing. Nvidia makes a ton of money there, and doesn't have to compete with AMD nearly as much as in graphics.

    DPFlops/Card seem to matter more. I suspect that DPFlops/Card matters most (due to IO issues) than DPFlops/$ (once known as machoflops, mostly for govt/academic epeenwaving).

    Now that both companies appear to be designing for GPU computing, it will be interesting to see how they compare (even if the 7970 seems to be missing half of its DPFlops. I wonder if they managed to de-power the transistors if they are there).
  • chizow - Tuesday, January 10, 2012 - link

    DPFLops/Watt discussions have a place, but not for desktop GPU parts branded GeForce or Radeon. If this were a Tesla part it'd be more meaningful.

    Nvidia cripples DP performance pretty badly on their GeForce parts starting with Fermi and I imagine they will do the same for Kepler. It also sounds like AMD is doing the same for Tahiti, but it will be some time before we have any idea if GCN is even directly competitive in real world GPU compute applications.

    Benchmarks and synthetic tests look great, but its going to take quite a bit of effort for AMD to get any penetration in a HPC market Nvidia has clearly dominated. Nvidia basically had to create their own API and GPU compute market from scratch, so AMD has their work cut out for them to catch up.
  • Jediron - Wednesday, January 11, 2012 - link

    No, not the Champions of the poor. But yes, Champions of the single core Videocard!
    GTX580, you lose.

    Sure the ball lies in Nvidia camp now, no suprise there! The ball was in AMD's camp, and they scored; that is obvious.

    Long live the red camp :-)
  • mhampton - Monday, January 9, 2012 - link

    Page 2 of the article starts by listing the setup tested, and the CPU must be a misprint - as far as I know there is no "i7-3936". Presumably this should be 3960 or 3930 instead.
  • know of fence - Monday, January 9, 2012 - link

    The same 3936 misprint can be found here:
    http://www.anandtech.com/show/5261/amd-radeon-hd-7...
  • Ryan Smith - Monday, January 9, 2012 - link

    Noted and fixed. Thank you.
  • BenSides - Monday, January 9, 2012 - link

    I agree with a poster above, who is apparently content running a single 5850 paired with the latest game releases. I have 2 5870's in crossfireX, in conjuction with an old i7 950 cpu. My benchmark results (Crysis) blow this new thing (7870) to hell.

    While not a gamer, benchmarking is a hobby of mine. Looking at the results here at Anandtech, it is reassuring to know that with the elderly cards I have, it seems there is nothing, at the present time, or looming on the horizon, which would make my graphics cards obsolete.

    In short, until AMD or NVIDIA introduce some revolutionary new technology, you're fine with your current card(s), if the presented benchmarks are any indication. Seems to me that at the present time, NVIDIA and AMD are producing new graphics solutions which are just overclocked versions of *old* graphics solutions.

Log in

Don't have an account? Sign up now