NVIDIA launch a new flagship Fermi: GTX580

NVIDIA GTX580 Performance Comparisons

While we'd have preferred to actually have a card to test ourselves, neither NVIDIA nor partner cards had arrived before the writing of this article. Generally we don't bother uploading these official graphs, and prefer our own more reliable testing. However there is some worthy analysis to be gleaned from the official NVIDIA performance slides included in the press presentation - but we won't be simply posting them all as given.

You're probably familiar with graphs such as the one below, which both AMD and NVIDIA - and indeed most of the industry - are guilty of. The graph shows a seemingly-phenomenal performance increase of one product against another, making the competing product (in this case the current high-end AMD 5870 and strangely the mid-range 6870) seem awful in comparison.

But there's one thing that is usually forgotten when looking at these graphs: the scale. As seen to the left of the graph, it doesn't begin at zero; rather beginning at 0.8. While NVIDIA was correct to standardise the results against the 5870, they didn't get the scale right - so we thought we'd fix their graphs for them. This is the untouched graph, showing an obliteration of the competition:

The untouched NVIDIA GTX580 performance comparison slide. Makes AMD look pretty bad!

And here's our version of the same information:

And Atomic's fixed performance graph. It still makes AMD look bad, but much less so than before.

Quite a difference, isn't there? We'll let those graphs speak for themselves (our own testing has shown that the GTX480 is more-or-less equal to the 5870 in far more tests than that graph shows). But where does all this performance improvement come from, and what part of the graph should be looked at? We think the most telling comparison is to ignore the red data for now, and focus instead on the NVIDIA - the difference between those green lines isn't very much. NVIDIA helpfully pulled that information out for us in the presentation.

Arch comparison between Fermi GF100 (GTX480) and GF110 (GTX580).

As seen in the above graph, NVIDIA reckon that at the same clocks, the GTX580 is faster than the GTX480. Called officially "Architectural Enhancements" rather than a new architecture, a boosted CUDA Core count of almost 7% accounts for most of the performance improvements seen in this graph. As for the rest, Jason explained that the GTX580 can now do “half-speed to full-speed FP16 calculation", and has "improvements in Z-cull algorithm” that give it the extra kick needed. Note that this is the same FP16 technology that NVIDIA slammed AMD over some months ago.

How does it go in DirectX 11 games?

Predictably, the NVIDIA GTX580 is a bit of a whiz kid when it comes to DirectX 11. Thanks to the hardware support for tessellation, it excels in situations when it is heavily used - which you can see to the right of this graph below. Interestingly the Civilization 5 shows a significant advantage; probably attributed to faster DirectCompute.

The GTX580 is predictably impressive in tessellation-heavy titles.

And finally, NVIDIA are touting great scaling under SLI, showing performance of up to two times better than AMD. It definitely looks impressive; but again, we'll have to wait until we actually have cards in our hands to confirm or reject their claims.

When you've got one fast GTX580, why not add another?

Read on for other improvements to the GTX580, including the heatsink and power system.