Memory is slow, Disk is fast – Part 1
12 days ago
- #computer-science
- #performance-optimization
- #hardware-trends
- Hardware improvements focus on wider capabilities (more cores, bandwidth, vector units) rather than faster clocks or IPC.
- Traditional metrics like clock speed, IPC, and memory latency have stagnated for decades.
- Memory bandwidth, caches, PCIe, and storage/network bandwidth continue to scale exponentially.
- Vector operations and parallel computing (e.g., AI workloads) benefit most from modern hardware trends.
- Single-threaded, non-vectorized software sees minimal performance gains due to stagnant fundamentals.
- Old assumptions (e.g., 'memory is faster than disk') are no longer universally true due to shifting hardware dynamics.
- Physics limits (e.g., quantum effects) hinder further improvements in latency and clock speeds.
- Traditional software must adapt to leverage parallelization and vectorization for future performance gains.
- Computer science principles (e.g., Big O notation) may need reevaluation given modern hardware realities.