nVidia's chip is the first to adopt the unified programming model for Microsoft's DirectX 10 software interface for graphics to ship with Windows Vista in January. Analysts give nVidia kudos on the design, but said they will wait to see competing parts, expected in the next few months from archrivals ATI Technologies and Intel Corp., before they render a final judgment.

"This [unified architecture] will be a major shift in PC graphics," said Dean McCarron, principal of market watcher Mercury Research (Cave Creek, Ariz.) "The nVidia part looks strong on paper, but we need to wait until all the products are out and running to determine who is ahead," he added.

Previously, graphics chips used separate arrays of custom hardware units optimized to handle tasks such as calculating vertex points or shading pixels. Sometimes chips bogged down when an application needed more resources for a particular process than a chip offered. Under the new architecture, data will stream across a generic pool of arithmetic logic units (ALUs) that can handle any graphics process needed.

ATI and Intel parts are expected to have roughly similar unified hardware architectures for their DX10 devices. "Implementations will vary, but if you have a DX10 part it will likely have this basic capability," said McCarron.

nVidia is taking that model one step farther by creating a special mode where those ALUs can handle computing threads instead of graphics primitives. That opens the door to easier use of the chips in data-intensive technical applications such as medical imaging or oil and gas exploration. Those apps traditionally use banks of DSPs, FPGAs or custom vector processors, but are increasingly turning to highly parallel graphics processors

"During the [four-year] process of designing this product it became clear some people were increasingly using graphics to do nongraphics applications. But that's quite awkward because you have to pretend to write graphics code when you are not, so the thought was to do this more directly," said David Kirk, chief scientist of nVidia.

ATI has given signs it sees a similar opportunity. The company demonstrated use of its graphics chips for "streaming" applications such as grid computing recently, although it may not disclose its next-generation core until early next year.

Currently, nVidia ships "a few thousand" graphics systems a year to users in a wide variety of technical computing markets. It hopes to expand that base when workstation and rack-mounted systems using the GeForce 8800 ship starting in January.

In its technical computing mode, the chip provides a thread scheduler in hardware and a facility for sharing cached data between ALUs. It lets programmers work in C rather than relatively arcane graphics languages currently used in scientific computing with existing graphics processors. At 681 million transistors the 90nm GeForce 8800 "is one of the largest semiconductors ever fabricated," said Kirk." The biggest challenge was getting the performance increase we wanted without astronomical increases in power," he added.

The GTX8800 adapter card provides twice the performance of nVidia's current graphics adapter. It consumes 145 watts maximum and 116 W on average. That compares to 116 W max and 99 W average on the last-generation card.

The GeForce 8800 organizes its 128 ALUs into eight processing units with 16-kbyte caches per block. The 32-bit, single-precision scalar processors have their own integer and floating-point units and a full instruction set. Versions supporting double-precision math and additional ALUs are in the works.

"This is a very scalable design," said Andy Keane, general manager of nVidia's technical computing group.

The chip handles its own geometry and physics processing, offloading the tasks often run on a CPU or separate accelerator, respectively. The result is greater smoothness of lines due to better anti-aliasing and superior rendering of highly complex images, especially ones involving shadows, smoke, skin and hair.

The chip also supports new levels of high-definition video post-processing, both for the H.261 and VC-1 codecs.

The GeForce 8800 comes in two flavors of 16x PCI Express adapter cards. The 8800GTX sports a 575-MHz core, 1,350-MHz shader block and up to 768 Mbytes of GDDR3 running at 900 MHz. It sells for $599.

The 8800GTS sells for $499. It includes 96 ALUs, a 500-MHz core, a 1,200-MHz shader and up to 640 Mbytes of GDDR3 running at 800 MHz.

Separately, nVidia announced upgrades of its chip sets for Advanced Micro Devices and Intel computers.

The 680a SLI for AMD processors supports dual graphics processors, a whopping 12 serial SATA interfaces, four Gbit Ethernet MACs and 20 USB 2.0 connections for high-end workstations. AMD is designing a reference board for the chip.

With the merger of AMD and archrival ATI, nVidia seems to be putting more energy into its Intel chip sets. The 680i SLI aims to enable aggressive overclocking for games enthusiasts using Intel's latest dual- and quad-core processors.

The chip set sports a new DDR2 memory controller and nVidia's overclocking Bios. It can goose a 1.8-GHz Intel Conroe processor up to 3.5 GHz, pushing its 1,066-MHz front side bus to 2,100 MHz. It can also step up memory from the traditional 800 MHz to 1,200 MHz.

That effectively lets users buy a $150 Conroe processor and run it at performance levels that rival a $1,000 Intel Extreme Edition CPU. "This is why people like to go overclocked. You can buy a cheap processor and get a lot out of it," said Matt Wuebbling, an nVidia chip set product manager.

nVidia has designed its own motherboard to showcase the features of the chip set.