AMD Cayman info (or rumor)

[zerazax]

I'm going to say this again: stop looking at synthetics!

First of all:

6950 - 1408 SPs * 800 * 2 = 2.252 TFlops
6970 - 1536 SPs * 880 * 2 = 2.703 TFlops

Now compare this to the other cards:
5850 - 1440 SPs * 725 * 2 = 2.088 TFlops
6870 - 1120 SPs * 900 * 2 = 2.016 TFlops
5870 - 1600 SPs * 850 * 2 = 2.720 TFlops

So in terms of theoretical shader power, Cayman and Cypress are a wash.

Why is this important? Well, look at how 5VLIW is implemented in Cypress and Barts:


For those who don't get the Cayman change to 4VLIW, it goes like this: R600,RV670,RV770, Cypress, and Barts all use the above image's shader configuration. Each SIMD has 4 simple shaders and 1 "fat" complex shader - the transcendental one - that can do complex math (such as calculating pi, sine/cosine, etc.). The other four shaders only do simple MADD (multiply-add). Since most shader operations are just MADD, the 5th shader often sits idle and thus 5-VLIW wasn't as efficient.

The move to 4-VLIW is this: instead of 4 simple + 1 complex, there's 4 moderately complex shaders. Each can still do MADD. However, to do a transcendental, I'm not sure how AMD is going to handle it - it might have one of the 4 shaders iteratively calculate a value, or it might use all 4 to do it at the same time so it requires fewer clock cycles.

Now why is that important? Because synthetics can utilize the 5th shader because they're designed to run em! So, synthetics often got to utilize the 5800's / 6800's up to their full shader potential, which is actually more or less a wash with the 6900's.

However, supposedly real gaming wasn't the case - most vector operations are shorter than 4 values, and so the 5th shader in Cypress/Barts is under-utilized.

Hence, in synthetics, the fact that Cypress is close to Cayman is a reflection of the fact that shading power in Cayman isn't vastly improved for benchmarks. The real key will be how gaming performance goes, since supposedly in games, going to 4-VLIW SIMD brings it to 98% the same performance as a 5-VLIW SIMD. The leaked gaming benchmarks seem to suggest that gaming performance is much higher than 3dMark is hinting at.

This can all be proven wrong of course, but it's the best reason for the gap between 3dMark/synthetic scores (to say nothing of Nvidia's optimizations in the tesselation benches) and gaming scores that were leaked

Also, people forget that the 5800's got HUGE boosts in 3dMark Vantage through drivers. In fact, a few boosted it scores by a considerable few hundreds of points. Who's to say that AMD doesn't do the same and pushes it upwards? The Extreme score on the 6970 already puts it above the 570, so we'll see where it goes. And if this slide is accurate, we still have drivers to use:




But that's why its pissing me off the guy won't use his GTX 480 to bench against the 6970 in games, and will only bench useless synthetics