AMD claims its Stream processor based on the year-old ATI R580 core has roughly similar single-precision performance of about 360 GFlops to the Nvidia GeForce 8800 launched last week.

The Stream CPU uses 48 pixel shaders running at 600 MHz with a 650 MHz memory bus. By contrast the GeForce 8800 sports 128 generic arithmetic logic units running at up to 1,350 MHz with a memory interconnect running up to 900 MHz. The Stream chip does not use any of its eight vertex shaders, a relic of its graphics heritage.

The Stream processor comes on a 16x PCI Express board with 1 Gbyte GDDR3 memory. It costs about $2,600 and dissipates 165W on average. Third parties including Rackable Systems are expected to deliver rack-mounted servers using the chip.

AMD is taking a slightly different software tack to Nvidia. It is making available a free license to the so-called Close To Metal (CTM) software interface for the graphics chip. The interface will let software developers tailor applications to the chips hardware resources. About 60 companies and academic groups are using the CTM interface which has been available in a beta version since July.

No native C compilers exist for the CTM interface, but they could appear over the next 12-18 months, said Dinesh Sharma, director of enterprise stream computing at AMD. By contrast, Nvidia said its 8800 will work in its technical computing mode with existing C compilers.

Sharma said most applications for the Stream CPU will be developed in house by large vertical market companies. One large oil and gas company already has its own application using Stream.

The ATI graphics division of AMD is expected to launch a fully unified graphics architecture before May that will adopt many of the concepts use in the Nvidia 8800.

"Its no big secret the future of graphics processors is in unified architecture" said Sharma.

For example, the Xenos graphics chip ATI helped design for the Microsoft Xbox 360 adopts a unified architecture with a pool of generic computing elements that can dynamically handle either pixel or vertex operations.