TypeScript 6 vs 7 on type-fest

Verdict

On an Apple M3 Pro (12 cores), TypeScript 7.0.2 checked the type-fest sources 3.5× faster than TypeScript 6.0.3 (0.484 s vs 1.67 s): 2.6× comes from the native compiler alone (single-threaded) and 1.3× from parallel checking. On the full suite with type tests, TypeScript 6 stops at Node.js's default heap limit (4.0 GiB); with a 16 GiB heap it finishes in 74.6 s, against 14.1 s for TypeScript 7 (5.3× faster), but TypeScript 7 peaks higher in memory (11.6 GiB RSS vs 10.4 GiB). Stays open until a second machine is measured with the same protocol.

Takeaway

Most of the gain is the native compiler, so expect it even on CI runners with few cores; more checkers help less (0.642 s with 1 checker, 0.377 s with 8). On large type-test suites, check peak memory before switching: TypeScript 7 needed no heap flag but used more memory at its peak.

Note Types and Instantiations differ between TypeScript 6 and 7 on identical code, so never compare these counters across compilers. TypeScript 7's counters also vary slightly from run to run on the full suite, single-threaded included; on the library sources they are identical across runs.

Note Memory is peak resident memory (RSS) from /usr/bin/time -l, field "maximum resident set size", of the largest process: the Node.js process for TypeScript 6, the native compiler for TypeScript 7 (its launcher replaces itself with the native binary). It is not the compiler's own "Memory used" line, which for TypeScript 6 is the V8 heap.

Note Out of memory means Node.js's default heap limit (4.0 GiB here), not the machine's RAM: TypeScript 6 aborted with "Reached heap limit" at 4.2 GiB RSS, and completed with --max-old-space-size=16384.

Note One library (type-fest at one commit), check time only: no emit, watch mode or editor latency.

Note The two machines differ by CPU as well as core count, so compare conditions within a machine; the comparison between machines is indicative only. On the 1-core Intel Xeon, parallel checking gives 1.0× (0.9× to 1.1×) and the native compiler 4.0×.

Note The Intel Xeon results use the earlier protocol: 3 runs, no warm-up, not interleaved, no RSS. The Apple M3 Pro results use a discarded warm-up per condition and interleaved runs, with other applications open (load average 3.0 to 5.0). The full-suite runs stay noisy: up to 6.7% coefficient of variation.

Note Checker counts were measured in a separate series from the default and single-threaded runs; compare them with each other.