What kind of memory bandwidth should we expect on 5800? I'm doing D3D11 dev primarily (OpenCL has crippling perf problems for me) and I'm only seeing 42GB/s read/write. This is for a shader of 512 threads where each thread reads 8 DWORDs in sequence from a structured UAV. Is there some better access pattern to use? Is it bad to write to a UAV you just read from - should I try ping ponging? The test array is 16M elements and I'm running the loop 10000 times for that 42GB/s figure.
