Conclusion

NVIDIA has painted to us a beast, a really big beast, as far as GF100 goes. The new GPU, on paper, is poised to keep NVIDIA's idea of leadership with design, complexity, and concept-execution alive, as it did with the G80 (GeForce 8800 GTX) and GT200 (GeForce GTX 285). It has restructured the GPU using logical sub-GPUs (the GPCs), which are tiny GPUs in themselves, and which add modularity and scalability to the design. To put it very coarsely, the GF100 is a "quad-core" GPU as far as giving hierarchy to raster operations go, and "many cores" at a lower pixel level. With tessellation, and how NVIDIA chose to implement it on the GF100, it looks like geometric complexity in consumer 3D graphics got the quantum leap which was overdue, which shading got many times over, meanwhile.

With 512 CUDA cores, the GF100 has over twice the SIMD horsepower the GT200 has, and can handle complex shader-intensive applications, with enough room left to handle other GPGPU tasks that run through the game such as PhysX (physics simulation), AI processing, and even more advanced post-processing technologies such as depth-of-field. The GF100 supports all major GPGPU interfaces, including CUDA, DirectCompute, and OpenCL, giving developers the room to code more than just graphics for this GPU. The 384-bit GDDR5 memory interface supplies the GPU with 50% more memory bandwidth, and 50% more memory size. Ideally consumer variants could feature 1.5 to 3 GB of memory on the board, with 192 GB/s of memory bandwidth (assuming the memory runs at 1 GHz). With present high-performance memory chips in the market, the bandwidth could end up much higher than that. Image quality enhancements are abundant, apart from new AA modes, the GF100 can work to reduce jittering, noise in shadows, and also work towards minimizing the performance penalty of these on a game's overall performance.

A lot is left to be demonstrated by NVIDIA. Such presentations work towards making the media and consumers aware of what the GPU is poised to be, what the company has developed, and what is going to be on offer, but they don't quite demonstrate the product itself. To the media the presentation lays out exactly the areas that need extensive testing and demonstration to prove they work as presented here.

The introduction of graphics cards based on the GF100, assuming it lives up to its expectations, will alter the high-end graphics market. Currently AMD has taken the lead with its powerful DirectX 11 GPUs to date, and has also released a graphics card that uses two of those GPUs (making for the most powerful graphics card money can buy). Prices may rise and fall, today's expensive GPU could become tomorrow's affordable one, or stay expensive, depending on how GF100 based graphics cards perform against market rivals in the ATI Radeon HD 5870 and Radeon HD 5970.

Aside from the fact that high-end graphics hardware are also very expensive to make and to sell, and then stay in the market to fetch profits, they also serve another vital purpose, to sell the company to buyers of smaller, more affordable graphics cards based on the architecture. This makes getting the performance lead, and retaining it doubly important for both AMD and NVIDIA. For NVIDIA, merely making the fastest GPU isn't going to work this time around, because it does not have any immediate mainstream or value derivatives of the GF100 as yet. AMD on the other hand, has been able to quickly put its performance lead to work, in selling mainstream and value GPUs. The company has made DirectX 11 GPUs for every market segment, $50 thru $700. A recent report also shows that the company has already shipped over 2 million of these GPUs. The number includes mobile DirectX 11 GPUs for notebooks, a segment where NVIDIA has not even announced plans to release any DX11 parts. So NVIDIA better get to work on GF100 derivatives, if it hasn't already. Lest it ends up with only $500 graphics cards for the better part of the year which cater to a very small portion of the market.

In all, here's an NVIDIA GPU with its task cut out: Perform.



Edit: NVIDIA has released a whitepaper on the GF100 architecture, which can be downloaded here.