345Explicit Parallelization Using MathWorks Toolboxes
3. GPUs can handle only simple arithmetic/indexing, and few toolbox functions.
4. GPUs can run code branching (if/else, logical indexing) but slower than CPUs —
GPUs work much better when there is no code branching.
5. GPUs work better with 100K-1M threads, not less (due to overhead costs).
6. GPUs do not work well with memory-intensive operations; so, they are better for
simple independent element-wise operations, but not to others.
7. GPUs are more memory limited than CPUs, although modern GPUs constantly
increase the amount of onboard RAM.
8. CPU parallelization of implicitly parallelized code (vectorized linear algebra, etc.)
is pointless; only GPU parallelization might help to ...