3DLabs' P10 Visual Processing Unit - When a CPU & GPU Collide

The P10's "Pixel Shaders"

One of the biggest rocks that 3DLabs is throwing at the ATI/NVIDIA camp is calling their "pixel shaders" not truly programmable. Instead, 3DLabs consistently refers to ATI/NVIDIA pixel shaders as register combiners that allow flexibility but aren't truly programmable. This goes back to the Carmack quote from earlier expressing displeasure with the way NVIDIA's pixel shaders turned out.

3DLabs doesn't discredit the power of these register combiners but they say it's not enough and thus the P10 offers a separate solution. The P10's programmable texture processor supports all current DX8 Pixel Shader specifications but, just like their vertex engine, it is composed of a number of 32-bit processors in parallel.

The P10's Four Pixel Pipelines
(Note: Yellow blocks are programmable, orange are fixed funtion)

The P10 has a total of 64 floating point coordinate generator processors that help determine texture coverage while another 64 integer processors are used to calculate the final pixel colors. As you can expect, all of these processors are fully programmable and as we just mentioned they support a superset of all currently available DX8 Pixel Shader functions.

It is important to note that although the P10 does offer a superset of all DX8 Pixel Shaders, this isn't a true DX9 part. One of the requirements for DX9 are floating-point pixel pipelines from start to finish, and as we just mentioned this simply isn't the case for the P10. Moving to floating point processors throughout the pipeline simply isn't economically viable for 3DLabs with the P10 when you take into account the increase in gates over the present mixed fp/integer setup. 3DLabs hypothesizes that they will need to be on at least a 0.13-micron process and maybe even a 0.1-micron process before full DX9 compliance will be possible. TSMC's 0.13-micron process isn't mature enough for 3DLabs to feel comfortable with from an economic standpoint, so we'll have to wait until next year for DX9 to be brought to this platform. You should also keep in mind that NVIDIA's next-generation part (NV30) will be made on TSMC's 0.13-micron process and will most likely be a full DX9 part when it comes out this fall.

In terms of other features supported by the P10's programmable texture processor, the VPU can apply 8 textures in a single pass (not a single clock) in comparison to the Radeon 8500's 6 and the GeForce4's 4. The processor also supports arbitrary texture sizes which will be important for Longhorn but again that is more of an enablement feature than something presently demanded.

Although common texture filtering algorithms are hard-wired into the pixel pipelines, the P10 also allows for custom filters to be programmed (e.g. custom anisotropic filtering shapes).

Another Programmable Stage

The final stage in the P10's pipeline is also made programmable which is something that no currently available GPU offers. After a pixel has been textured and filtered then other techniques such as anti-aliasing can kick in and do their thing. In all other GPUs these final processes are not programmable and are all hard-wired into the pipeline, but the P10 they are not.

The P10 supports virtually any type of anti-aliasing you can think of; edge anti-aliasing is supported under OpenGL which is the AA method of choice as the performance hit is significantly less than more brute force methods while offering excellent line AA quality. Supersampling and multisampling are supported as well; the former offers virtually unlimited samples while you are limited to a maximum of 8x when multisampling.

Because of the final programmable pixel stage the P10 can support much higher precision color depths such as 64-bit color. A feature you'll end up seeing more of going forward is the support for gamma-correct 10-bit RGB outputs which is built into the P10. The VPU also has 10-bit DACs to take advantage of the custom 10:10:10:2 (RGBA) mode.