3 ms·
I agree, but very little of this work seems exciting. While advances like hardware support for transactional memory and, hopefully, some unified memory support
by rotskoff 13y ago
I agree, but very little of this work seems exciting. While advances like hardware support for transactional memory and, hopefully, some unified memory support for CPU/GPU operations will have a tremendous impact on the ease with which programmers compute in parallel, I tend to hear fewer new ideas about what can be done with parallel computing.
- kjs3 13y agoExciting is subjective. :-) I don't disagree that there are fewer ideas, but that's because we need new ideas, including new architectures. Having spent quality time on a number of novel supercomputer architectures that seem to have been forgotten in the rush to turn GPUs into short vector SIMD-ish successors to Cray XMP/YMPs, I don't think we've looked at all the viable alternatives, at least not in a long time. For example, I'm clear on why memory-based vectors like the Cyber-205 faded at the time, but it's not clear the same limitations fundamentally still exist (especially with L2/L3 caches bigger than the 205 main store now). I think that's interesting...YMMV.