Nvidia's next-gen GPU arrives in March to challenge Radeon 5800 series

Nvidia's next-generation GeForce GF100 (aka Fermi) GPUs for Windows 7 desktops will finally be available starting March 26. Nvidia will launch its first DirectX 11 GPUs, the GeForce GTX 480 and GTX 470, at a gaming convention in Boston.

Nvidia decided to tip its hand earlier this week after a couple moves generated confusion. First, on Monday Nvidia announced three GeForce GT 300 series GPUs. The company had billed this as a "major announcement," but these turned out to be renamed versions of existing GT 200 series GPUs, and available only in new PCs, not in retail add-in-boards. Many hardware enthusiasts and tech journalists had assumed that the Fermi GPUs would be the GeForce GT 300 series. Then listings for GeForce GTX 480 boards started popping up on a retail site (don't try ordering one).

The extent to which the GF100 has been delayed is a matter of debate, but it has certainly had a long gestation period. Nvidia first discussed Fermi at its GPU Technology Conference in September 2009, where the company emphasized features tailored more for high-performance computing than for gaming. At the Consumer Electronics Show in January, Nvidia demonstrated what looked like production-level boards on a handful of DirectX 11 games and benchmarks, but still provided few details.

Part of the reason the GF100 has taken so long, as Nvidia engineers have hinted, is that designing a large, complex GPU that is equally suited to both general-purpose computing and 3D gaming isn't easy (just ask Intel). Fermi is an ambitious undertaking with up to 512 cores and a total of around 2 billion transistors But the GF100 has also been hindered by manufacturing issues. Both AMD and Nvidia use the same semiconductor foundry, Taiwan Semiconductor Manufacturing Company, to manufacture their most advanced GPUs on the latest 40nm technology. TSMC has admitted to trouble getting its 40nm process up and running.

This has clearly affected both AMD and Nvidia. AMD has had lots of time to work out the kinks with 40nm technology. The company launched its first 40nm GPU, the Radeon HD 4770, in April 2008. A couple months later, AMD and TSMC showed off the first 40nm DirectX 11 GPUs, and in September AMD released the ATI Radeon 5800 series. Since then AMD has filled out the line and during the holidays it sold 2 million DirectX 11 GPUs, but executives said they could've sold more were it not for manufacturing glitches. This week a TSMC executive said the 40nm manufacturing issues had been resolved, and the foundry planned to double its capacity on that node this year. Separately Nvidia CFO David White said the manufacturing issues would be fully resolved around the middle of the year when Fermi-based boards will be available in volume.

Because of its size and complexity--along with the 40nm manufacturing challenges--there's some question whether Nvidia will be able to sell these Fermi boards in any large numbers. I'll leave that debate to others. My only interest is what kind of numbers a GeForce GTX 480 can put up. This isn't just an issue for gamers. An increasing number of consumer and professional applications are tapping into the GPU to speed-up tasks such as video encoding that can be handled by lots of cores in parallel. At this point, no one outside Nvidia really knows. My guess: the GeForce GTX 480 will do enough to allow Nvidia to claim the fastest single GPU, compared to the Radeon 5870, though it's not likely to beat out the Radeon 5970, which puts two of the Cypress 5800 series GPUs on a single board. There's plenty of room for Nvidia to split the difference--graphics cards based on the Radeon 5870 currently start at around $400, while the dual-GPU 5970s with 2GB of graphics memory start around $650.

In about a month we'll know just what kind of fight Fermi has in it. I look forward to testing out the first GTX 480 boards.