Performance

This page compares Mightty with Ghostty 1.4.0 and kitty 0.49.2 on start-up time, memory, throughput, CPU use and latency, and explains how each figure was measured.

How it was measured

The figures come from bench/suite.py in the Mightty repository, run on an M2 with macOS 15.

• Each terminal is launched fresh at 120×40 in Menlo 14, with every other setting at its default.
• A probe runs inside the terminal, and the suite samples the terminal's process every 50 ms.
• The terminals take turns, round by round, so that a busy moment on the machine doesn't fall on only one of them.
• Each figure is the median of 10 launches or 3 sessions.

Absolute numbers change from run to run with the machine and its displays. Compare the terminals within a row, not against figures from another run.

Start-up

Memory

Memory here is the physical footprint that macOS reports, which is the figure Activity Monitor shows as Memory.

The scrollback defaults differ: Mightty keeps 10,000 lines, Ghostty keeps 10 MB and kitty keeps 2,000 lines. When a pane's output pauses, Mightty releases the memory its scrollback lines hold but don't use.

Throughput

Each terminal prints a 16 MiB file with cat, timed from inside the terminal.

Single runs vary by about 10% on a busy machine, so a difference of a few percent between Mightty and Ghostty is not meaningful.

CPU

100% is one CPU core. At a steady 250 lines a second, Mightty and kitty use about the same amount.

Replying to programs

Programs ask the terminal things, such as where the cursor is, and wait for the answer. This is the time from asking to having the answer:

Mightty reads a pane's output in pieces of at most 1 KiB and sends any reply after each piece. A reply doesn't wait for the screen to be drawn.

Typing

This is the time from a key press to the frame that shows its echo, on a 60 Hz display. Only Mightty is measured, because the probe sends keys through the app's own event queue and reads the time Metal reports for the frame reaching the screen, which isn't possible in another terminal. The figures don't include the keyboard or the window server, which come before the app.

After a key press, Mightty draws about 2 ms later, once the echo has arrived, instead of waiting for the next refresh. With vsync on, that frame still waits for the display, behind any frame already queued. With vsync off, the display takes the newest frame, at the risk of tearing. vsync is on by default, as it is in Ghostty and kitty.

Second throughput run

bench/compare.sh measures six workloads, alternating the terminals over 5 rounds of 3 runs. It runs Mightty as a pane in a copy of Mightty that is already open, not in a freshly launched one, so these figures are not measured the same way as the ones above and are from an earlier run.

For long lines, Mightty and Ghostty both reach the rate at which macOS can pass output to a terminal. Pooled over 63 runs, their medians are 258 and 263 MiB/s.

Running the benchmarks

python3 -I bench/suite.py         # everything above except the second run
bench/compare.sh                  # the second run; needs Mightty open

See Architecture for how output gets from a program to the screen.