Jeff & Sanjay published an updated version of the famous “Latency Numbers Every Programmer Should Know” from 2007. Below is a comparison of the two:
| Operation | 2007 | 2025 | Speedup |
|---|---|---|---|
| L1 cache reference | 0.5 ns | 0.5 ns | 1× |
| L2 cache reference | 7 ns | 3 ns | 2.3× |
| Branch mispredict | 5 ns | 5 ns | 1× |
| Mutex lock / unlock | 100 ns | 15 ns | 6.7× |
| Main memory reference | 100 ns | 50 ns | 2× |
| Compress 1 KB | 10 µs | 1 µs | 10× |
| Read 4 KB (storage) | — | 20 µs (SSD) | — |
| Read 1 MB over network | 10 ms (1 Gbps) | 0.1 ms (100 Gbps) | 100× |
| Read 1 MB from memory | 250 µs | 64 µs | 3.9× |
| Datacenter round trip | 500 µs | 50 µs | 10× |
| Disk seek | 10 ms | 5 ms | 2× |
| Read 1 MB sequentially (disk) | 30 ms | 10 ms | 3× |
| CA ↔ Netherlands ↔ CA RTT | 150 ms | 150 ms | 1× |
Sources: