Memory guide, part 2

The CPU handles program logic (such as artificial intelligence decisions for games), the "geometry preprocessing" (object culling and dismantling curved surfaces into polygons), and more. Sending those polygons to the video chipset is the only bandwidth intensive operation that has to be done. In other words, only 3D-applications which use huge amounts of polygons (100,000 per scene and more) will demand high amounts of bandwidth. Therefore, while a decent amount of bandwidth is necessary for "geometry streams," cache misses are still the primary reason why the CPU accesses main memory. Cache misses imply that the CPU is impatiently waiting for a certain piece of data. So the latency, or the numbers of clock cycles required for the "critical word" to arrive, will determine the system's efficiency and overall performance. Low latency memory will perform better in most applications than high bandwidth solutions for our typical set of desktop applications.

Bandwidth, the ability to burst huge amounts of data to the CPU, will become more important as games will use higher polygon counts and take advantage of more 3DNow! and iSSE code. iSSE and 3DNow! all include prefetch commands. Prefetch commands fetch a piece of data a certain program will need within a few tens of clock cycles. For example, consider a prefetch operation that runs inside a loop, requesting data needed for the next iteration. In the case of an array of data being processed by the loop, each successive element could be requested in advance. In these terms, prefetching can be thought of as a sort of "non-faulting load." In ideal circumstances, data and instructions will, thanks to prefetching, always be in the L1-cache when the CPU needs them. Latency is far less important in this case since the CPU never waits for a piece of data it is dependent upon, and bandwidth will determine the performance of the system.