According to the Ethash algorithm described here, the computation for each nonce requires a loop of 64 iterations, where each iteration reads a chunk of 128 bytes from a pseudo-random location in memory.
Reported hash rates for the NVIDIA GTX 1070 are around 30MH/s, which means that every second this GPU is reading 30.000.000 nonces x 64 iterations x 128 bytes from memory. That is 245.76 GB/s.
I was unable to find the peak memory bandwidth that the 1070 is capable of, but this thread suggests it should be 197.76 GB/s, and Ethash is achieving even more than that. Howcome?!
Also, I would expect that reading from many different memory locations (as opposed to sequential reads) would result in a very inefficient use of memory, but it seems it doesn't really matter?
I'd appreciate if someone could clear this up, thank you!