What I wanted to say is ith the GPU, memory access have a known latency (you are writing to the L1 equivalent) without a statistical component. Also the nature

What I wanted to say is ith the GPU, memory access have a known latency (you are writing to the L1 equivalent) without a statistical component. Also the nature of tasks submitted on the GPU (reified) makes it easier to optimize imho.