> "now each machine needs a couple of pricey GPUs."
Well it would be much more pricey to get the same amount of flops out of traditional CPUs.
Well it would be much more pricey to get the same amount of flops out of traditional CPUs.
The current approach of both CUDA and OpenCL camps to move to some sort of standard intermediate representation is a step in the right direction.
This might open the door to more declarative approaches for data modelling in the GPU.
[1] Or OpenCL, though I don't know anyone that uses it...too slow.
Up to recently OpenCL was all about C. Plus it followed the same broken model as GLSL shaders, requiring low level compiler as library.
I think OpenCL camp is mostly unaware how advanced CUDA is in regard to language support, since they live in pure C land.
Better performance is a consequence of NVidia cards being quite good in regards to the competition and they pushing their own tech.
Overall you are quite right. As I mentioned, I see the move to PTX, HSAIL and SPIR as a small step in that direction.
More needs to happen in terms of the whole architecture.
Old habits die hard, but for the past 20 years, the official spelling has been: Fortran.