Sweeney comments on HLSL, GPU architectures
Beyond3D has some interesting commentary from Epic's Tim Sweeney on the limitations of Microsoft's High Level Shading Language (HLSL) and how he'd like to see GPU architectures change:
But there's only so much useful work you can do in a non-Turing-complete programming language like HLSL lacking persistent memory reads and writes, synchronization primitives, and the ability to call arbitrary functions at runtime (i.e. via function pointers on a variable-sized call stack). In other words, I don't think there will be a big usage case for 2500-instruction or 25,000-instruction shader programs that perform nothing but integer and floating-point arithmetic and texture sampling.
Sweeney thinks that GPU architectures will have to change significantly for the current pace of eye candy improvement to continue. His comments are short, but definitely worth checking out