3 ms·
Sufficiently smart compilers never showed up where it mattered most for Intel: Turning branchy C/C++ system and apps code into explicitly parallel instructions
by jacques_chester 12y ago
Sufficiently smart compilers never showed up where it mattered most for Intel:
Turning branchy C/C++ system and apps code into explicitly parallel instructions.
And arguably, they can't.
Sufficiently smart compilers for GPUs aren't sufficiently smart. They work by imposing sufficiently dumb limits on programmers instead.
- fiatmoney 12y agoInterestingly, Intel spends a lot of die budget on on-chip instruction level parallelism, and their major instruction set-level enhancements have been increasing vector operation lengths. They've also invested somewhat heavily in GPGPU-style vector coprocessors in their Xeon Phi's.
- jacques_chester 12y agoRight. But the whole point of EPIC was to do away with that part of the die budget entirely, and a lot more besides. GPGPU programming has worked because nobody proposed to run general code on it from the beginning. It just so happens that there are some workloads that fit the vector processing model. But lots still don't. Lots and lots and lots.