View Single Post
Old 01-24-12, 03:09 PM   #136
Registered User
Join Date: Jul 2003
Posts: 1,526
Default Re: next gen kepler to support dx 11.1, also take a year to rollout all cards

Originally Posted by ninelven View Post
No, unless you are using AA, the max for GF110 is 24.83 Gigapixels/sec. Tahiti is 29.6 GP/sec by the same math.

No. It doesn't.

The numbers are 9.75 for Fermi and 13.3 for Tahiti. Why isn't Tahiti getting 29.6? Because the test is bandwidth limited. Now, the 7970 has 37% more bandwidth than the 580, so if the test is bandwidth limited we should expect the 7970 to perform 37% better than the 580. 13.3 / 9.75 = 1.364. SHOCKING. The 7970's advantage over the 580 is exactly equal to its bandwidth advantage in a bandwidth limited test... imagine that.

This is the third time I have explained the result, and I really don't think I can make it any simpler or more clear. If at this point you can not understand that the test is completely bandwidth limited, then I must conclude you are either 1) intentionally trolling and not worth having a conversation with or 2) genuinely intellectually incapable of understanding and not worth having a conversation with.

The test wasn't using AA at all, so that doesn't come into it to begin with, and i fully know that the test is bandwith limited, but it does to show that theoretical maximums mean nothing if it can't be delivered in real terms.

To make matters worse, you stated that AA use doesn't drop the theoretical maximum fillrate on the HD7970, while it does on the GF110 to 24 Gigapixels so score one more point for AMD there, and above all else, those 48 rops on the GF110 might actually take up more room on the die than the 32 used in the HD7970.

It's about efficiency in the end so adding more of them in kepler is pointless as it just takes up more die space and there isn't enough memory bandwith to feed them all anyhow....Add fewer and make them more flexible in not dropping fillrate when AA is used like you mentioned for the HD7970.
shadow001 is offline   Reply With Quote