3 ms·
In my experience, it's not bandwidth that is the limiting factor, but latency. You'll hit the same problem with FPGAs if you're using it as a co-processor, as t
by scott_s 15y ago
In my experience, it's not bandwidth that is the limiting factor, but latency. You'll hit the same problem with FPGAs if you're using it as a co-processor, as they are typically connected to the motherboard over PCI Express. If the vectors you're using are small (where "small" means small enough to easily fit into an L1 cache on a processor), then you probably won't see any performance improvement by offloading the computation to an accelerator.
I say this because in a matrix-vector multiplication, only the vector has data-reuse. You do a single pass over the matrix. I wrote a paper where latency killed any performance benefit from using a GPU, because the computation we performed did only a single pass over the data: http://people.cs.vt.edu/~scschnei/papers/debs2010.pdf http://people.cs.vt.edu/~scschnei/papers/debs2010.pdf If you're doing a matrix-matrix multiplication, then that's a different story because each element in each matrix will be reused.