P10: GPU Becomes VPU
It has taken almost two years for 3Dlabs to develop the P10. According to 3Dlabs, many of the established technologies were thrown out, and a fully overhauled chip design was born.
The chip has more than 70 million transistors, which even surpasses the Pentium 4 on this count. The costs for the research and development in this project ran up to nearly $22 million. The P10 is produced via the 0.15 micron process, and 0.13 micron is being prepared. 3Dlabs'attention is focused on the full programmability of pixel and vertex shaders - thus the name Visual Processing Unit (VPU). Even textures do not necessarily have to be loaded, and instead, they can be programmed. 3Dlabs keeps the memory management completely virtual.
The P10 can logically address 16 GB. However, Wildcat VP cards will only be available on the market with 64 and 128 MB memory, the rest has to be tapped from slow system memory if needed. The P10 uses 256 instructions per program to process the vertices (transform, lighting and clipping). A maximum of 22 opcodes, 256 constant registers and 16 work registers can be used for this purpose.
The new chip from 3Dlabs has four parallel pipelines for textures. With respect to vectors, they go one step further. Here, scalars can be used, instead of the traditional vectors. The advantage of the scalar approach is that a larger number of vertices can be processed simultaneously.