also the code tends to be highly SIMD, code with a lot of data dependent branches benefits very little from higher memory bandwidth when the fetch latency from ram is on the order of hundreds or even thousands of cycles.