The slides from Tim Sweeney's High Performance Graphics 2009 keynote expand upon his Next Mainstream Programming Language talk we discussed some year back. To summarse, Tim makes the following points (in my interpretation):

• The current GPU pipeline is too limiting and too difficult to program


• The CPU roadmap from Intel and others show a convergence between the CPU and GPU. We're going to see a lot of cores with vector units hanging off a shared cache. The cores are getting simpler but programming them is getting harder


• Programs that are purely functional can be automatically vectorised, and the new CPUs are going to extend the scope of vectorisation to general programs


• Economic reality demands tools to make programming this hardware easier. New languages seem the only viable way forward

It's an interesting talk with a bit more detail at the bit-bashing end than previous talk. The question is: who is going to make this language?