Nvidia Demos Next-Gen Fermi Graphics

"Fermi is a radically different way of designing GPUs," Jen-Hsun said following his keynote to open the GPU Technology Conference. Nvidia hosted a similar, more gaming-oriented event called Nvision last year but changed the name of the conference and tightened the focus on GPU-based parallel and cloud computing this time around.

Nvidia has not revealed a shipping date or pricing for Fermi-based products, which would succeed the GeForce GTX 280 and other products based on the Santa Clara, Calif.-based chip maker's current, 10th-generation G80 microarchitecture. But Jen-Hsun noted in a press conference that with working Fermi silicon in Nvidia's hands, "time to market is traditionally just a few short months."

Nvidia partners exhibiting products at the conference could not provide firm shipping dates for Fermi, but several suggested April as a likely time frame. That would put Nvidia as much as two quarters behind rival Advanced Micro Devices, which late last month debuted its next-generation ATI Radeon 5800 series products, regarded by many reviewers as the most powerful GPUs currently on the market.

"Obviously, you'd like to have a product to ship when the other guy has a new product," Jen-Hsun conceded Wednesday, but argued that Nvidia's focus is less on battling for product market share these days and more on developing the ecosystem for its CUDA programming language for GPUs and "unified computing" that exploits the benefits of both CPUs and GPUs for faster, more powerful computing performance.

The Fermi architecture will be deployed throughout Nvidia's three major product lines, GeForce, Quadro and Tesla, Jen-Hsun said. The Fermi chip contains 3 billion transistors -- about three times as many as current-generation Nvidia GPUs -- and will be the first Nvidia product to support DirectX 11, which AMD's recently released GPUs support today.

"They're just transistors. You shouldn't get too attached to them," Jen-Hsun joked when asked about the size of the Fermi die, which is expected to be the largest ever for a company that already builds the world's biggest volume-manufactured chips. Fermi will reportedly organize those 3 billion transistors into 512 CUDA processing cores, that are in turn arranged as 16 streaming multiprocessors with 32 cores apiece.

In the demo, Jen-Hsun showed "brand-new" Fermi silicon performing a visually rendered exercise in "peak double-precision arithmetic," used in several critical high-performance computing (HPC) applications. The next-generation GPU does that particular bit of computing about eight times better than the current generation of Nvidia processors, Jen-Hsun said.

Attendees at the conference agreed that Fermi will make its most dramatic first showings in the supercomputer and HPC markets. Jen-Hsun himself hinted at a future Fermi-based supercomputer to be built by the Oak Ridge National Laboratory possibly rocketing to a very high position on the Top 500 Supercomputers list.