DRAM is the new I/O. So yes, this is designed to handle 97% of the threads doing constant bad-locality DRAM accesses.
The protocols for HPC are so amorphous that they bubbled up into the lowest common denominator, completely software defined async global workspace