Smaller grid size per GPU does not speed up the calculation significantly about bluebottle-3.0

[ajsierakowski]

Are you comparing to a single-threaded CPU code? Multi-threaded? Distributed? The only reason I can think of why a code would run more than twice as fast when cutting the problem size in half (or, said another way, that scales less than linearly with problem size) is if you are seeing significant communication bottlenecks that decrease in influence as problem size decreases. In the single-GPU case, inter-process communication is negligible, which is why it scales very nearly linearly with problem size.

[tomchen95]

I'm actually just considering computations on just 1 GPU, so there's no communication consideration. If the global computational domain size is 160160160 and 16016080 respectively, 160160160 should run at least 2 times slower than 16016080, provided that only 1 GPU is used for both cases. But based on my test, 160^3 is not that slow and 16016080 is not that fast. So I'm wondering why.