Karl Berg, on 26 December 2014 - 12:21 PM, said:
Exceedingly cool! Thank you very much for the links sir! I will definitely be reading through them in detail.
I've followed GPGPU work at a high level for some time, as it really seems to be the closest one can achieve to Cell / SPU style programming on a PC. I'm really excited to one day see that first GPU vendor really nail down an SPU style interface, with full application level control over DMA channels, a working C/C++ compiler for program binaries, and full intrinsics for assembler support, as well as 128-bit or better SIMD vector support. I initially thought AMD might provide such an interface when I first read about their new APU tech, but they seem to have taken a different direction. Regardless, once someone opens raw metal access with a decent API, I think the sheer number of FLOPs those GPU chips are capable of will simply explode in terms of high profile applications.
For us on the server side specifically, since our code is so integrated with CryEngine and CryPhysics, it would be a significant challenge to re-implement our HSR systems to use GPGPU algorithms. Being able to write external helper programs with explicit DMA control, as described above would mitigate that pain a great deal, since there would be potential for sharing code and potentially even reusing some of the PS3 job scheduling logic already built into the engine.
Client side, there are some super easy wins with GPU based particle simulation. Our version of CryEngine doesn't support GPU based particle simulation; and we haven't had the resources to attempt this rewrite ourselves yet.
This is almost complete gibberish to me, but I love the enthousiasm I'm detecting here!