Beginning with the publication of my mid-October, 2005 print and website series 'Instigating A Platform Tug of War', I've closely followed GPUs' expansion beyond 2- and 3-D graphics acceleration into more general-purpose functions. A creative tension exists between GPGPU advocates' desires for expanded chip capabilities (such as on-board instruction cache), tailored to broader potential application (audio, still imaging and video DSP, physics processing, database sorting, etc) needs, versus the unrelenting silicon cost-minimization pressure of the still-dominant graphics function.

I suggested, and in a later article on gaming consoles demonstrated, that one possible technique for addressing both camps' desires was to migrate GPUs beyond dedicated pixel and vertex shaders and towards a unified shader architecture that could therefore also address the processing of other types of data sets. Specifically, I showcased the unified shader-based Xenos GPU that ATI Technologies developed for Microsoft's Xbox 360 console. While I pointed out that Nvidia's competitive GPU in the Sony PlayStation 3 was of the traditional vertex-and-pixel shader variety, I suggested that Nvidia's response to ATI's unified shader challenge wouldn't be long in arriving.

Nvidia confirmed my prediction with the mid-November unveiling of the world's first DirectX 10-compliant GPUs, the GeForce 8800 GTS and GTX, which respectively cost $449 (640 MByte frame buffer) and $599 (768 MByte) MSRP in board form. Instead of giving a long litany of speeds and feeds, I'll showcase some notable data points below, and direct you to the FAQ, technical specification and additional documentation pages on Nvidia's website for more details:

As I mentioned in my 2006 WinHEC coverage, it'll likely be quite some time before a critical mass of DX10-based applications emerges. But  the GeForce 8800 series will be equally relevant in the current DX9-pervasive era. Published benchmark results back up Nvidia's website claims that a single GeForce 8800 GTX GPU is faster than dual GeForce 7900 GTX GPUs running in SLI mode, as well as faster than a single GeForce 7950 GX2. For other perspectives on Nvidia's latest achievement and the broader-picture GPGPU opportunity, see the following chronologically-ordered writeups:

• Nvidia Launches 8800 Series, First of the DirectX 10 Cards
• NVIDIA's GeForce 8800 (G80): GPUs Re-architected for DirectX 10
• NVIDIA rethinks the GPU with the new GeForce 8800
• GPUs to Power Supercomputing's Next Revolution
• GeForce 8800 Roundup: The Best of the Best

My GeForce 8800 GTX reference board is in-hand and, assuming I can track down a sufficiently beefy power supply (and AC outlet), I plan to test-drive it in my next reference system, now slowly-but-surely under assembly. Stay tuned.

What's next for Nvidia? In the briefing the company gave me on its C-based CUDA GPGPU development tools suite, it indicated that whereas the 8800 series GPUs supported single-precision floating point math, double precision-capable GPUs were under development. And what about ATI? The company had a relatively quiet 2006; the same Microsoft-induced distraction that temporarily sidelined Nvidia seems to have afflicted it, and the company's acquisition by AMD has likely further negatively impacted productivity. But I strongly suspect that ATI's DX10-compatible GPUs will soon appear, no later than the end-of-month Windows Vista launch and perhaps as soon as the upcoming CES (greetings from Las Vegas, I've just arrived). And I equally strongly suspect that ATI's chips will be of the unified shader flavour, following in the footsteps blazed both by the company's Xbox 360 GPU and competitor Nvidia's offerings.

Digg This | Slashdot This