Conclusions
I didn't expect Nvidia to produce a unified shader architecture for this generation of product, and I certainly wasn't anticipating anything quite like this GPU. Honestly, I just didn't foresee the company taking such a huge risk with the GeForce FX debacle still relatively fresh in its mind. As Nvidia CEO Jen-Hsun Huang pointed out at the G80 press event, designing a more traditional graphics pipeline in silicon leads to fairly predictable performance characteristics. The performance of a graphics processor that depends on stream processors and thread-based load balancing is much more difficult to model. You don't necessarily know what you're going to have until it's in silicon. The history of graphics is littered with failed chip designs that attempted to implement a non-traditional pipeline or to make a big leap toward more general programmability.

Fortunately, the green team did take that risk, and they managed to pull it off. I still can't believe the G80 is a collection of scalar stream processors, and I'm shocked that it performs so well. The real-world, delivered performance in today's OpenGL and DirectX 9-class games is roughly twice that of the Radeon X1950 XTX. Yes, it's taken 680 million transistors to get there, but this kind of performance gain from one generation to the next is remarkable, a testament to the effectiveness of the G80's implementation of unified shaders.

The G80 has just about everything else one could ask of a new GPU architecture, too. The new features and innovation are legion, anchored by the push for compliance with DirectX 10 and its new capabilities. The G71's texture filtering problems have been banished, and the G80 sets a new standard for image quality in terms of both texture filtering and edge antialiasing. This GPU's texture filtering hardware at last—or once again—delivers angle-independent anisotropic filtering at its default settings, and coverage sampled antialiasing offers the feathery-smooth quality of 16X sample sizes without being a major drag on frame rates. Despite being large enough to scare the cattle, the G80's doesn't draw much more power under load than the Radeon X1950 XTX. The chip is still too large and consumes too much power at idle, but this architecture should be a sweetheart once it makes the transition to a 65nm fab process, which is where it really belongs.

No doubt Nvidia will derive an entire family of products from this basic technology, scaling it back to meet the needs of lower price points. Those family members should be arriving soon, too, given the way DirectX 10 deprecates prior generations of graphics chips. Huang has said Nvidia is taping out "about a chip a month" right now. I'm curious to see how they scale this design down to smaller footprints, given that the ROP partitions are tied to memory paths. Will we see a mid-range graphics card with a 192-bit interface? I'm also intrigued by the possibilities for broader parallelism in new configurations. Will we see arrays of G80-derived GPUs sitting behind one of those external display chips?

Our time with the G80 so far has been limited, and as a result, I've regrettably not scratched the surface in several areas, including the importance of DirectX 10 and CUDA, Nvidia's analog to ATI's stream computing initiative. We will have to address those topics in more detail in the near future. They are related in that both DX10 and CUDA expose new ways to harness the power of the G80's stream processors for non-traditional uses, either for graphics-related tasks like geometry shaders that can create and destroy vertices or for non-graphics apps like gaming physics and scientific computing. Now that the GPU has cast off even more of its fixed pipeline and gained more general stream processing capacity, the applications for its power are much broader. Both Microsoft and Nvidia are working to give developers the tools to take advantage of that power—and soon.

We also haven't tested the GeForce 8800 in SLI. Although Nvidia says SLI will be an option for consumers starting today, they actively discouraged us from testing SLI and didn't provide us with an SLI-capable driver. Fair enough—I didn't even do justice to DX10 and CUDA, and besides, who needs two of these things right now? Folks who want to do SLI with an 8800 GTX will need to have a power supply with four PCIe aux power plugs—or some converters and lots of open four-pin Molex plugs. They'll also probably need to have a 30" LCD display capable of 2560x1600 resolution in order to use this magnitude of GPU power in current games. Regardless, we are interested in the new possibilities for things like SLI AA, so we will be exploring GeForce 8800 SLI as soon as we can.

For now, the G80—excuse me, GeForce 8800—is the new, unquestioned king of the hill all by itself, no second graphics card needed.