On most Linux distributions and when the file is mmap'd, if you run it a second time the data will still be in RAM and not have to be read from disk. This gets the runtime down to 1.1s for this AMD 2950x (https://github.com/gunnarmorling/1brc/discussions/46#discuss...).

With SIMD and certain assumptions about the input this can seemingly be further reduced to well under a second, eg see https://github.com/gunnarmorling/1brc/discussions/138.