jmke

The evidence of this is in the performance of SuperPi 1M, seen as a benchmark that largely fits in 2MB cache and doesn’t depend heavily on bandwidth. Run four copies, though, and each takes 20% longer than if two copies had been run on a Conroe of the same speed: 25.3 seconds compared to 21 seconds. So what’s causing this problem? Our theory is that a strength of the core, the intelligent prefetchers, are getting in each other’s way, saturating the bus, and increasing the latency to RAM. Move to a more bandwidth-hungry application and the problem will surely just get worse

__________________