Inside Fermi: Nvidia's HPC Push

To some extent Fermi looks like more of the same, only bigger and faster.

What's really interesting about it is that they've added L1 and L2 cache, they've gone to a global memory model so you can pass pointers around and they've put some effort into making branches less expensive. I didn't yet spot whether it supports recursion, but it's implied that it does. I expect to see a lot more GPU-raytracing projects start popping up in the near future thanks to this...

I'd be very interested to see how this compares to Larrabee, if/when it actually comes out.

There's a pile more information on nvidia's site for fermi at http://www.nvidia.com/object/fermi_architecture.html

The big change is certainly the support for indirect jump instructions and the exception handling support. Quite a big leap in GPU programmability. The only piece still missing is fully coherent caches.

Single page: http://www.realworldtech.com/includes/templates/articles.cfm...

Good find.