3Dlabs' Visual Processing Architecture implements an optimized graphics pipeline, replacing previously inflexible pipeline stages with highly programmable SIMD (single instruction, multiple data) processor arrays. The P10 VPU combines over 200 SIMD processors throughout its geometry, texture and pixel processing pipeline stages to deliver over 170Gflops and one TeraOp of programmable graphics performance together with a full 256-bit DDR memory interface for up to 20GBytes/sec of memory bandwidth.Hmm.. What was that I was saying? This thing has a virtual memory subsystem and a multithreaded command processor. (Read the whole press release for all the dirt.) 3Dlabs calls it a VPU, or Visual Processing Unit, just to be contrarian.
3Dlabs' Visual Processing Architecture avoids the increasingly difficult-to-program collection of arbitrary functional units found in today's more limited GPUs. Instead, 3Dlabs provides application developers a clean, orthogonal, compiler-friendly architecture that implements many CPU-like architectural innovations.
Obviously, this sounds like a killer chip. 3Dlabs is promising rock-solid drivers for DX9 and OpenGL 2.0, and they're expecting cards to hit the market in the third quarter of this year. With twice the memory bandwidth of a GF4 Ti 4600, you can bet they won't be slow. (Thanks to Kert Chian for the tip.)