AMD's Catalyst Evolution For The Radeon HD 7000 Series

[smitty3268]

I hope you realize that the kind of execution unit does not matter for stream processing, only the bitrate matters. Kepler has 3.5Tflops@64bitFMAC or 7Tflops@32bitFMAC, wile Radeon7000 has 4Tflops@32bitfmac. Second, there is not any possibility for some one to produce a GPU that does not fill the shaders because of memory for example, that way where is the proof? (why the card reaches maximum thermal and use?).

/headdesk

I give up, I don't think there's anyway i can convince you of your mistakes here.

I defy you to present even a single app ever created by AMD or NVidia that has ever reached the full theoretical performance of a piece of hardware they produced. Just one.

Try taking a look at overclocking your memory versus overclocking the cores. Notice that 1 will make more of a difference? Notice that it will often depend on the app you are using as to which one matters more? Notice that effects are often completely different on other models?

[barkas]

/headdesk

I give up, I don't think there's anyway i can convince you of your mistakes here.

I defy you to present even a single app ever created by AMD or NVidia that has ever reached the full theoretical performance of a piece of hardware they produced. Just one.

Try taking a look at overclocking your memory versus overclocking the cores. Notice that 1 will make more of a difference? Notice that it will often depend on the app you are using as to which one matters more? Notice that effects are often completely different on other models?

Those numbers look wrong, too, since kepler has a 3:1 ratio of 32bit to 64bit. Indeed I found 3TFLOPS @ 32bit for kepler on the net. Indeed, Performance figures for the Tesla K10 card, which is clocked lower than the normal graphics parts, are 2290GFlops/SP and 95GFlops/DP per chip. So the Desktop part probably does about 3TFlops/SP and 150GFlops/DP.

@artivision: care to name your sources? Because I can't find anything in the same ballpark even.