AMD, Nvidia Processors Go Head To Head

Nvidia and Advanced Micro Devices (AMD) recently released new graphics processors as the two chip makers continued to compete on discrete graphics products while emerging as allies in promoting the possibilities of GPU computing in the face of CPU giant Intel's uncertain steps in that direction.

Nvidia's new GeForce GTX 200 series includes the GeForce GTX 280, available for a suggested manufacturers' price of $649, and the GeForce GTX 260, with a price tag of $399. The GTX 280, with 240 processors and a full gigabyte of frame buffer memory, is clearly Santa Clara, Calif.-based Nvidia's latest edge-pushing, high-end consumer card, while the 192-processor, 896-MB GTX 260 has humbler specs, but not by much.

Up the road in Sunnyvale, Calif., AMD's latest pair of consumer cards from its ATI graphics division targets a lower portion of the discrete market. The new ATI Radeon HD 4800 series will deliver a teraflop of graphics performance, according to Rick Bergman, GM of AMD's Graphics Products Group.

And the Radeon HD 4850, released in late June, will do it for less than $200, Bergman said. The 4870, released in early July, is more powerful than the 4850, and more expensive at roughly $300.

"We've adopted a new model for this product and and#91;we'reand#93; going forward," Bergman said, outlining what he described as an important shift in AMD's strategy for discrete graphics. "It's targeting the $200 to $300 part of the market because we think that's the meat of the market."

The third card in the 4800 series, aimed at the $500-range ultra-enthusiast segment, is the ATI Radeon HD 4870 X2, initially priced at $549. Released Aug. 12 in tandem with the more affordable Radeon HD 4850 X2 ($399), the 4870 X2 is described by AMD as the world's fastest graphics card, delivering 2.4 teraflops of processing power on a single card. It's also the first consumer graphics card with two gigabytes of memory, according to AMD.

Some analysts believe the 4870 X2, coupled with some Nvidia price cuts, marks the return of head-to-head price competition between ATI and Nvidia at the high-end.

Next: Tesla Vs. FireStream

Tesla Vs. FireStream
One area where Nvidia and AMD seem to be taking each other on more squarely is the world of general-purpose GPU computing on high-performance compute clusters. Graphics-only Nvidia has been more vocal about the potential of GP-GPU computing in recent months, devoting an entire Technology Editor's Day in June to its Tesla specialty cards and CUDA language for parallel programming for GPU clusters.

In addition to the consumer products, both vendors had second-generation, teraflop-performing treats for HPC partners. Nvidia in early summer unveiled its $1,699 Tesla 10P GPU, while AMD countered with its $999 FireStream 9250 card for GP-GPU computing on single-socket servers and workstations.

This space may still be somewhat esoteric, but with Nvidia pushing hard for CUDA while AMD touts FireStream and an open-standards approach to parallel programming, it's certainly competitive. Yet much as it is on the consumer client side, the Nvidia-AMD competition for HPC customers is tempered by a mutual interest in confronting the leviathan looming over all things silicon.

That would be Intel Corp., Santa Clara, Calif., which has rather famously dismissed GP-GPU computing and discrete graphics leadership as necessary to its immediate business, though the new message around its forthcoming Larrabee chips have modified that position a good deal. Intel, has also made it clear that its take on graphics will be significantly different than current technology models.

Intel CTO Justin Rattner reiterated the message at Research@Intel Day, telling the crowd that Intel sees ray tracing as the future of visual computing, set to replace traditional raster-based graphics in the years to come. Intel, which has been cryptic about the next-generation graphics technology code-named Larrabee, released a paper at August's Siggraph conference outlining some details about the chips and its overall graphics ambitions. Surprisingly, the paper was fairly brief in its discussion of ray tracing, instead focusing on the benefits of bringing the x86 architecture to graphics computing, as Intel plans to do by late 2009 at the earliest.

Whatever Larrabee turns out to be, Nvidia and AMD can at least take heart that it might not even arrive until 2010. That's a mighty long window for the two companies to continue playing the part of "frenemies" when it comes to jointly evangelizing discrete graphicsand#8212;while fighting tooth and nail in the product trenches.