Casey Destroys Optimization Myths | TheStandup

Casey debunks a viral tweet claiming that replacing division with reciprocal multiplication always improves performance. Key points: (1) floating-point division and multiplication are not mathematically equivalent due to precision loss, which can matter in scientific computing; (2) modern CPUs (e.g., Zen 4/5) handle floating-point division far faster than the tweet's claimed 20–40 cycles — closer to 3–10 cycles latency, with throughput of 2–3 cycles vs. 0.5 for multiply; (3) in a load-op-store loop, the bottleneck is usually memory bandwidth, not the arithmetic operation itself, making the divide-vs-multiply distinction largely irrelevant; (4) blindly applying such micro-optimizations without profiling and understanding the full pipeline leads to wasted effort and incorrect mental models about performance.