It's interesting to learn that major motion picture studios apply 128-bit precision to create rich, realistic computer-generated scenes. By matching these film industry levels of precision, the NVIDIA GeForce FX GPUs enable high-quality images with spectacular cinematic effects, without any artifacts or compromises in quality, and the real-time application of those effects throughout the entire scene.

The keywords for GeForce FX are:

• 0,13 micron GPU fabrication
• 125 million transistors (give or take a few, I didn't count)
• DDR2 memory clocked at 1 GHz
• 51 billion floating point operations per second (51 gigaflops) in the pixel shader alone
• Advanced Programmability (3rd generation)
• High-precision color (64-bit & 128-bit color)
• High-level Shading Language
• New Vertex and Pixel Shading instructions
• Highly efficient architecture (3rd generation Lightspeed Memory Architecture)
• High Bandwidth to memory and CPU
• Shaders can be 1000's of instructions long.
• 8 pixels per clockcycle rendering power
• 200 Million Triangles per second
• 64-bit & 128-bit color, this is film-quality precision, in fact a higher precision than the movie Toy Story 2 used. 64-bit offers high precision with 2x the performance & half the memory of 128-bit. As it seems developers want both 64 and 128-bit color precision for advanced effects.
• AGP 8x (over 2GB/sec bandwidth to the system).
• Fully DirectX9 compatible
• Pixel Shaders 2.0
• 0.13 Micron fabrication process.

One of the early NVIDIA documents we received about 8 weeks ago stated something interesting about performance, the GeForce FX would be able to do:

• 2x to 3x GeForce 4's frame rates
• 3x GeForce 4's vertex processing

NVIDIA even included some benchmarks in that initial document, check this out:

In all honesty the numbers seem to be very rough and theoretical, GeForce 4 (NV25) indicates 100% as 1, the rest you can figure out. If indeed the final performance will produce numbers like these then thit is extremely impressive.

The numbers will get even more interesting when I tell you that they are based on 1280x1024x32-bit colors at 4x FSAA. The NV25 in this case was a GeForce4 Ti 4600 and HQ Pixel Fill = Dual Anisotropic texture @ 32-bit colors.

The received document we received back then clearly stated that DX9 entails a strong shift from bandwidth towards computation. Basically the new bottleneck seems to be computing efficiency over memory efficiency. As stated above the NV30 GPU has 3rd generation LMA, NVIDIA states that entails a 1.0 GHz memory data rate but an internal 48GB/sec effective bandwidth due to LMA III, that remains speculative though.

Dawn technology demo ..