FP16 and FP32 Precision

Increased precision

One of the key features that the next generation graphics accelerators such as
GeForceFX and RADEON 9700 bring are their support for higher precision 16-bit and 32-bit floating point formats (64-bit and 128-bit color). For the ultimate visual quality, developers can utilize the 32-bit floating-point format, delivering the same level of precision currently used in the film industry today.

NVIDIA's Toys demo

All plastic models utilize a special plastic shader, while wood and brushed metal shaders are also present

Some of you may be wondering why the need for millions of colors, when no monitor is capable of supporting that range. As images become more complex, the 256 levels of each color today’s 32-bit accelerator’s are capable of generating aren’t enough to produce lifelike images. Rounding errors can occur, resulting in images with artifacts. A few situations where this occurs frequently are in bump maps, smoke, or fog.


Many of today’s 3D games use hacks to work around the limitations of lower precision. Another alternative that is often used is to perform these calculations on the CPU. As a result, these technologies limit what the programmer can do and lower performance.

Obviously rendering an entire scene in 128-bit color mode would require a considerable amount of horsepower (and thus kill frame rates if the hardware isn’t able to keep up), with this in mind GeForce FX also supports the 16-bit floating point format for situations when full 128-bit color isn’t required, and thus increasing performance.

Developers are free to move back and forth between both formats, using the format that is ideal for their application. For instance, high-resolution textures can utilize the 32-bit floating-point format to create objects with an unprecedented level of detail, other applications can be performed using the 16-bit floating-point format and the increased performance it offers.

Time machine demo

As the truck gets older, pixel shaders are used to render the truck's advanced age

Performance

During cinematic sequences game developers could use this to create highly detailed protagonists, while ordinary characters that aren’t crucial to the plot could be rendered with less detail. The end result is that developers can produce cinematic graphics in real time while optimizing the performance for every situation.

NVIDIA feels its GeForce FX GPU goes one step beyond RADEON 9700 PRO in the fact that it truly supports 128-bits of color. RADEON 9700 PRO is limited to just 96-bits of precision, which could prevent the RADEON 9700 PRO from matching GeForce FX’s 64-bit performance as the core may not be intelligent enough to split the data into 64-bits. It also won’t be able to match the visual quality of GeForce FX in 128-bit color mode.