On the other hand, memory speed (both latency and bandwidth) hasn't changed much in that timeframe. That means that relative to the speed of the CPU, it takes much longer to access memory now than a decade ago, and so stuff that made sense to cache in memory back then will probably be faster to recompute whenever needed nowadays instead.
back
1 comments
Maximum theoretical bandwidth may not have increased at the same pace but modern memory controllers are a lot more efficient. I'd wager that the CPUs are kept adequately fed even without extremely fast RAM like on GPUs.