The release of Longhorn will also be critical to both NVIDIA and ATI's plans - given the timing and the fact that there are not to be any low end NV5x, this may suggest that NV5x is set to be an extended Shader 3.0 architecture, which would indicate that the further Longhorn moves into 2007 the better it would suit NVIDIA's architectural innovation cycle. Presently the expectation is that ATI will introduce their Shader 3.0 part, suggested to be primarily developed by the R300 architectural team, in mid 2005The rumormongers at The Inq would seem to concur. They are also expecting the next-gen ATI, apparently dubbed R520 and "developed by the R300 team," to support Shader Model 3.0. Both publications are reporting that post-3.0 shader capabilities won't debut until the "Windows Graphics Foundation" arrives in Longhorn in 2006 or 2007.
Neither report speculates on how these developments might affect the next big expected architectural innovation in GPUs, the generalization of on-chip computational resources, so that vertex and pixel shaders could use the same silicon. If I had to guess, I'd say it won't happen in R520 and NV50 if they're limited to SM 3.0, but that's just a guess. Beyond that, the decision to stay with SM 3.0 shouldn't hinder real-time graphics much. The shader program lengths offered by SM 3.0 are well beyond the real-time capabilities of today's fastest graphics cards, so the programming model has lots of headroom.