Moore's Law's End? The Bottleneck of Traditional Software Performance

Over the past 20 years, certain aspects of hardware have advanced rapidly (e.g., core counts, bandwidth, vector units), but instructions per cycle, IPC, and latency have stagnated. This breaks old rules of thumb, such as "memory is faster than disk." The article argues that traditional software (single-threaded, non-vectorized) performance gains are limited by these stagnant metrics, leading to skyrocketing cache miss costs. The author suggests we need to rethink how we write software to fully utilize ever-improving hardware capabilities.