WITH LESS THAN three weeks to go before the expected launch of the - now not ATI but just AMD - Radeon HD6000 series, that is, Southern Islands GPUs, the question remains, what should Nvidia do?
Or rather, what can it do quickly. The Green Goblin's new architectures for later during 2011 are not the subject here, but what happens before this Christmas shopping season.
First, let's think about what AMD might come up with on 19 October - an HD6770 with overall performance akin to the current HD5850, and an HD6870 with possibly up to 30 per cent higher performance than the default HD5870, or about 15 per cent above the souped-up HD5870 overclocked editions. Then, the dual GPU version might follow a month later, still well in time for last minute Christmas geek sales.
Now, if the HD6770 really performs this well, it will outpace any Geforce GTX460 in the market, including the factory overclocked editions. In fact, even AMD's new midrange card will likely perform at a level equal to the default reference GTX470 designs.
But the HD6870 might cause even more of a stir. Even if it's just a quarter faster than the reference HD5870, it will be hitting benchmark scores above those of all GTX480 cards. And as we've seen many times before, a win at the top of the GPU sweepstakes has a 'waterfall' effect that benefits the lower-end product sales for the 'champion' vendor, too.
So, what can Nvidia do to counter this AMD attack? At the GTX460 level, enabling more shaders - up to 384 - on the GF104 die might help quite a bit as long as Nvidia doesn't lower the clock frequency to reduce the heat. If a GTX480 GF100 die has up to 512 shaders, then the three-quarters of that on the GTX460 should allow for up to 384 shaders.
Of course, Nvidia had no intention to enable those earlier, partly to keep the yields up and also not to cannibalise GTX470 sales. The extra 13 per cent or so speed-up from the extra shaders would keep the competition quite even with any HD6770 part.
The situation is more complex for Nvidia on its competitive approach to the expected HD6870. There is nothing to replace the GTX480 GF100 die until well into next year, so a combination of tweaks like a bit of extra clock speed and an improved memory controller, as well as enabling all 512 shader cores in selected GTX480 chips, that is, GTX485 or such editions, would make good sense.
Even if the combined advantages only bring, say 10 per cent extra real performance, it would help Nvidia stay close to AMD. But, unlike the GTX460 versus HD6770 case, the high end battle here would likely still be won by AMD whatever the case. Yet, the difference would be so small that Nvidia shouldn't have a major problem selling the updated GF100 to its faithful fans.
However, the same might not be the case with the dual-GPU twin GF104 card part, with the Radeon HD6970 expected sooner or later. The dual-GPU twin GF104 card will start to look less and less appealing, as a pair of GF104s will be no match for even an AMD reference dual-GPU part.
Nvidia will have to work out another approach for the top-end dual GPU market, but it probably won't be able to do that until its next generation 28nm chips are ready. µ