Let’s assume that we have two GPUs A and B where both have the same theoretical memory bandwidth but A has twice the memory clock speed as B and 1/2 the memory interface width. Can we pretty much assume that the effective memory bandwidth is the same in these two GPUs? If not, what would cause a difference in performance? Moreover, what exactly does it mean when the GPU has a x-bit interface? Does that mean that the max amount of data that can be sent from DRAM to fast memory is set at x bit (meaning that there would be latency if multiple SMs fetch data in overlapping clock cycles where amount of total data is larger than x bit)?
If I’m not mistaking, the max bandwidth is calculated as
bit-interface * memory-clock * 2 (Double Data Rate) / 8 (bits in a byte) B/s
I’m no expert on the effects of memory interfaces but I think the interface could effect the execution speed of a kernel depending on whether or not it is susceptible to partition camping.
e.g. The GTX 480 has six 64-bit memory partitions for a 384-bit memory interface while the GTX 470 features five 64-bit memory partitions for a 320-bit memory interface.