Conclusion

On paper, NVIDIA’s GF100 GPU looks very impressive. With over twice the stream processors and a 384-bit GDDR5 memory interface, the GPU certainly looks like it should be able to deliver two times the performance of its predecessor, GT200. And with its four raster engines and 16 PolyMorph engines, GF100 could offer blazing tessellation performance.

Generally if you can double your graphics performance from one generation to the next you’re in good shape, and if you can do better than that you’ll end up ruling that generation of GPUs.

ATI’s RV870 chip still isn’t offering 2X the performance of RV770: as we noted in our Radeon 5870 Performance Preview, ATI’s dual GPU Radeon 4870 X2 was faster than the 5870 in multiple cases, so ATI has left NVIDIA an opening here.

The question is, what kind of clocks will NVIDIA manage to settle on for GF100? Will they push the clocks to 10/10ths for max performance at the expense of yields? Or will they play it more conservative, say opting for 8/10ths, in order to improve their yields (and profits) per GPU? In our opinion, NVIDIA played it too safe with GeForce GTX 280’s speeds, they underestimated Radeon 4870 and ultimately ended up paying for it in the end.

NVIDIA needs to push the pedal to the metal with GF100’s clocks. Since they’re late to market, everyone’s expecting a part that will be worthy of the wait. At the same time though, they’ve got a 3.0 billion transistor chip that’s bigger than the competition, and will therefore likely suffer from lower yields. This may encourage NVIDIA to play it safe with the clock speeds.

It appears that NVIDIA realizes this. While they haven’t disclosed the final clocks, initial results with prototype GF100 boards look promising. We weren’t able to run any benchmarks, but NVIDIA did run some tests with Far Cry 2 and Dark Void. Using Far Cry 2’s built-in benchmark with settings of 1920x1200 with 4xAA/16xAF, GF100 delivered a frame rate of 84.3 frames per second. In comparison, GTX 285 ran at 50.4 fps. More importantly though, the minimum frame rates for GF100 were dramatically higher: 65 fps for GF100 versus 38.3 for GTX 285. In the Dark Void benchmark running at the same level of AA and resolution, GTX 285 ran at an average frame rate of 38.48 fps with a minimum frame rate of 30.57 fps. GF100 was twice as fast, running at 76.82 fps with a minimum frame rate of 49.9 fps.

Naturally we frown on controlled testing that’s out of our hands, especially when we’re only looking at two benchmarks, but if GF100 can ultimately deliver two times the performance of GTX 285 across a variety of games, that would be quite a coup for GF100.

Now the question is, how much will it cost, and when will it ultimately be available for the public to buy? The latest rumors suggest that GF100 won’t be launched until sometime in March, with actual availability possibly coming as late as April. If true, that’s certainly not good news for NVIDIA. That gives AMD plenty of time to concoct a faster Radeon 5870 derivative, if not improve the performance of their drivers. It’s going to be an interesting next two months for sure…