Picking the right unit reduces to dividing by 1024 until the value is small, which Math.log(bytes) / Math.log(1024) does in a single step. Flooring the log gives the unit index; clamping with Math.min prevents overflow past the largest unit you have a label for. IEC units (KiB, MiB) are technically correct for binary multiples and are increasingly common in developer tools (Linux du -h, GitHub release sizes). Use Math.abs(bytes) so negative values (rare, but seen in delta accounting) format the same way.

Disk vendors, cloud providers, and most consumer-facing UIs use the decimal (SI) convention: 1 KB = 1000 B, 1 MB = 1000 KB. The math is identical to the binary version with 1024 swapped for 1000. Pick this when you want display values that match what the user reads on a hardware label or storage console; pick the binary version when the audience is technical and the source is in-memory bytes (Node.js process.memoryUsage, JVM heap stats). Mixing the two in the same UI is the surest way to create distrust.

A single helper with options handles the most common variations: pick decimal vs binary, set the precision, and choose a separator (a regular space, a non-breaking space \u00A0 so the number and unit do not wrap, or even an empty string for tight CSS). Showing whole bytes without trailing zeros (512 B instead of 512.00 B) is a small touch that makes the output feel less robotic; the idx === 0 branch handles that. Wrap this in your design system's number formatter (Intl.NumberFormat for locale-aware decimals) when localisation matters.