February 28, 2026
Fast facts, faster fights
Latency numbers every programmer should know
The speed cheat sheet that launched a comment war over typos, proof, and GPUs
TLDR: A refreshed open-source speed chart for computer tasks ignited debate: a 5ns vs 5µs typo, demands for proof, and requests to add GPU and RDMA (fast memory sharing) stats. It matters because these tiny delays determine how snappy apps feel and help engineers avoid hidden slowdowns.
A retro-styled chart of tiny time delays—like how long a computer takes to grab memory or send a small message—just dropped, and the comments instantly turned into a stopwatch smackdown. The headline drama: “Send 2,000 bytes over commodity network: 5ns.” One eagle-eyed reader shot back, “Shouldn’t this be 5µs?”, sparking a trust spiral over whether the numbers are precise, current, or just legendary lore. Another demanded receipts: “How accurate are they now?” and asked for differences across Intel, AMD, Apple’s M1 chips, and even GPU memory speeds.
Power users rushed in with add-ons. One insisted the list include RDMA—a fast way for computers to share memory across a network—at “~2.5µs,” flexing nerd cred while raising the bar. A confused-but-delighted commenter admitted the figures didn’t mean much to them, yet called the site “an excellent idea,” capturing the “I’m lost but I love it” vibe. For extra context, someone dropped a helpful CPU cost infographic while others pointed to the open-source GitHub repo and the classic Jeff Dean list.
The mood? Respect for the cheat sheet, but no free passes: fix typos, show sources, and expand for modern hardware. Nanoseconds vs microseconds became the meme of the day—and a reminder that tiny numbers can cause big fights.
Key Points
- •The article is a console port of “Jeff Dean’s latency numbers,” hosted on GitHub (chubin/late.nz) under MIT License, and references colin-scott’s interactive_latencies.
- •CPU and memory figures include: L1 cache 1 ns, branch mispredict 3 ns, L2 cache 4 ns, mutex lock/unlock 16 ns, and main memory 100 ns.
- •Data movement/computation examples include reading 1,000,000 bytes sequentially from memory at 741 ns and compressing 1 KB with Snappy at 2.0 µs.
- •Storage latencies include SSD random read at 16.0 µs, SSD sequential read of 1,000,000 bytes at 12.245 µs, disk seek at 1.649384 ms, and disk sequential read of 1,000,000 bytes at 358.968 µs.
- •Network latencies listed include sending 2,000 bytes over a commodity network at 5 ns, a same-datacenter round trip at 500.0 µs, and a California–Netherlands packet round trip at 150.0 ms.