each gpu to gpu transaction goes through the NB so its card to card bandwidth and latency is directly effected by the cpu FSB (on intel)