AMD has launched a verbal assault on Nvidia over its Fermi architecture.
While Nvidia claims its delayed next-generation graphics standard is a significant "game-changer" in the history of the GPU, AMD says that's cobblers and the Fermi chips, when they finally appear, will be too big to be viable and too late to be competitive.
Nvidia was yesterday showing off Fermi behind closed doors to vetted geeky hacks, and we assume that was why its spokesmen have been too busy to answer our emails.
Elsewhere, AMD was fighting back by dishing the dirt on what it thinks Nvidia is up to. And dish it did, long into the night, until our ears were aching and our pencil-weilding fingers swollen.
AMD's graphics division is in confident mood right now. They still call themselves ATI and can claim to be well ahead of their arch rival, having now shown a full range of DirectX11 graphics products out and about, from mainstream to mobile."We have a complete line-up of DirectX11 cards from top to bottom. Nvidia is a long way off that," a spokesman bragged.
AMD says Nvidia is pushing the Cuda parallel-processing ability of its chips because it has lost the graphics race. "The early announcement is kinda unprecedented, or rather, it's uncommon to give architectural details of a GPU several months in advance of the chip being available," said David Nalasco AMD's senior technical marketing manager. He reckons Nvidia is concentrating the argument on architecture because it can't compete in raw 3D power.
"The result is you get something that is isn't necessarily laser-focused on providing the best gaming graphics experience to the broadest range of end users.
"We see this as a an offensive move to try to get people put off buying their competitor's product rather than promoting their own product," said Nalasco in an exclusive briefing with THINQ, aimed at stealing Nvidia's thunder. The irony may have been lost on him.
What AMD really wants to talk about is silicon. And the size of its chips compared to Nvidia's.
Making any sort of chip for the consumer market requires the ability to crank them out by the shed-load. The main determinant in the cost of volume - assuming you get the design right - is processor size. Smaller chips mean less waste, more volume. AMD has managed to transition its graphics chip output to 40nm. Nvidia hasn't. Not only does AMD have DX11 chips on sale today, it is also able to leverage greater economies of scale thanks to its use of a 40nm process.
Spookily, both AMD and ATI use the Taiwan Semiconductor Manufacturing Company (TSMC) chip foundry to fabricate their chips. It means they can keep a close eye on each other. ANd in comparison to AMD's 40nm offerings, Fermi is a big chip.
"The latest information from our sources is that it [the Fermi chip] is going to be about 5-600mm square, that would put it at about 70 percent larger than a Cypress chip," Nalasco claimed.
"You have a chip that is definitely going to be very powerful and it is going to support DX11, but if you look at just how much its going to cost to deliver, you see Nvidia has some challenges to overcome.
"This is going to impact not just when they initially release the chip but also in a couple of months time when they're going to have to start getting into versions of the chips that actually are mainstream.
"When you make a really large chip it reduces the amount of the wafer you can actually use. You may only end up with a handful of useful chips from a wafer."
The process size leaves NV with significant problems, in AMD's view. "If it performs really well and everyone wants to go out and buy one, there is no possible way they are going to be able to supply them, even if they were able to somehow create demand for this chip
"If you're Nvidia and trying to boost your sales, you're not going to spend your wafers on producing these big chips, you're going to spend your wafers on smaller chips which, in their case, is older-generation technology."
Since the graphics chips from both firms are still all manufactured by TSMC, any differences in the chips are all coming from design not differences in fabrication.
TSMC has to try to share that demand between customers, "primarily between us and Nvidia", said AMD. With its smaller process size, AMD can get significantly more chips out of each run at the line that Nvidia can.
So, according to AMD, if Fermi is really fast, "if it creates lots of demand, there is no possible way they going meet that demand. if performance is no so great those people that were waiting to see what NV's response to the 5000 series would be would go out and buy a radeon product."
"They're damned if they don't. Damned if they do," Nalasco said.