Skip to content

Performance

Ö. Efe D. edited this page Sep 15, 2026 · 1 revision

Performance

Speed is tron's first goal.

Benchmarks

cat of large files in a 100x30 window on Fedora, Wayland, with an RTX 2070 SUPER. Median of five runs; lower is better.

Workload tron Alacritty 0.17 Ghostty 1.3
Plain ASCII, 54 MiB 689 ms 868 ms 1173 ms
True color SGR, 44 MiB 611 ms 728 ms 2969 ms
Unicode and CJK, 31 MiB 453 ms 449 ms 632 ms
Launch until the shell runs 55 ms 138 ms 419 ms

Without rendering, the parser handles about 500 MiB/s of plain ASCII, 270 MiB/s of true color SGR and 200 MiB/s of Unicode on the same machine:

cargo run --release -p tron-core --example throughput

How tron stays fast

  • Parser. A table-free VT500 state machine hands runs of printable ASCII and UTF-8 to the screen in one call, found with SIMD byte scans. Common CSI sequences, such as colors and cursor movement, are parsed in a single pass.
  • Screen model. Cells are 16 bytes. Scrollback and screen share one ring buffer, so scrolling a full screen is O(1), and clearing a recycled row touches only the cells that were written.
  • Rendering. The pty reader parses without waiting for frames. Each frame copies only the changed rows under a short lock and builds the GPU data from that copy. Rows that scroll reuse the GPU data built for them earlier. Everything is drawn as instances of one quad.
  • Glyphs. Shaped runs and rasterized glyphs are cached; box drawing and block glyphs are generated, not loaded from fonts.
  • Startup. The GPU opens on a background thread while fonts load and the shell starts, and compiled pipelines are cached on disk.
  • Frame pacing. Frames are drawn only when something changed, and animated shaders run at the display's rate.

Measuring latency

TRON_TRACE_LATENCY=1 tron logs the time from each key press to the frame that shows its effect.

Clone this wiki locally