Re:Linux founder tells Intel to stop inventing magic instruc...
by PaulHoule on Monday July 13, 2020 @05:13PM
(#23825653)
CUDA is night-and-day different from Intel's nightmare.
CUDA in particular hides the exact width of the processing elements from the programmer, so you don't have to rewrite your CUDA applications to take advantage of the next generation of GPU the way you have to do with AVX1,2,3,...
Mind you, NVIDIA's drivers compile your shaders to their proprietary instruction set, which is a path not taken in conventional CPU architecture even though it keeps coming up. (e.g. the belt)
Re:Linux founder tells Intel to stop inventing magic instruc...
by dragandj on Monday July 13, 2020 @08:01PM
(#23827095)
FWIW, Intel supports OpenCL for their CPUs, so you can use similar general parallel programming code which is compiled down to AVX-512 (fairly) transparently.
CUDA in particular hides the exact width of the processing elements from the programmer, so you don't have to rewrite your CUDA applications to take advantage of the next generation of GPU the way you have to do with AVX1,2,3,...
Mind you, NVIDIA's drivers compile your shaders to their proprietary instruction set, which is a path not taken in conventional CPU architecture even though it keeps coming up. (e.g. the belt)