https://developer.nvidia.com/blog/how-access-global-memory-e...
SIMT still expects coalesced memory access that's close together otherwise performance falls off a cliff
SIMT still expects coalesced memory access that's close together otherwise performance falls off a cliff