What is needed to increase the bandwidth of memory systems? Would this not be more of a hardware improvement? The software developer can't physically increase the bandwidth limit themselves right? The most developers can do is ensure that their code is optimized to utilize the cache in a smart way to minimize transactions.

Would tiling the vectors and say using a multiword caching scheme that leverages spatial locality decrease the bandwidth requirement for this vector operations? One optimization, for matrices at least, involves tiling matrices to increase redundant memory access and take advantage of temporal locality thus eliminating the latency cost for repeated data elements; I was wondering if this could be applied here, but using spatial locality in place of temporal?