Performance

Latest valid measurements for the CEF 156 engine. Each result names its commit, host and workload. The full record, including earlier runs and limits, is in docs/benchmarks.md. Native client latency and memory are on the benchmarks page.

1080p120 recording on macOS

Measured 2026-10-23 at commit 4694aa1 on an Apple M4 (16 GB, macOS 27.0.1). IOSurface capture, Metal NV12 conversion and h264_videotoolbox MP4, with 48 kHz page audio, the per-frame canvas-reallocation fixture and five CPU load workers.

Run Distinct FPS Held paints Audio Load average
120 s 119.07 112 (0.78%) no silence 3.9–8.4
60 s 119.25 45 (0.62%) no silence 8.0–11.4
10 s, A/V sync 120.00 0 8.33 ms audio delay 9.6–10.2

The recorder dropped and coalesced no frames; held paints are frames CEF never delivered. The strict sustained 120 FPS gate did not pass on the longer runs on this busy host. An earlier run of the same command on 2026-10-20, before the runtime modularization and on a quieter host, measured 120.00 distinct FPS over 120 s with no held paints.

Reproduce

Terminal
python3 scripts/verify-native-macos-recording.py \
  --installation '/path/to/relocated installation' \
  --output /new/output/dir --format mp4 \
  --duration-ms 120000 --case 1920x1080@120:audio --load-workers 5 \
  --required-distinct-fps 120 --require-no-recorder-drops --require-native-capture \
  --require-source-continuity

Build and relocate the runtime with the macOS full-runtime guide first. Add --audio-sync-proof for the A/V sync row, and state the fixture mode with any number you publish.

Limits

macOS ARM64 only. Arbitrary page workloads, popup paths and live streaming throughput are not covered. A 120 FPS container can still repeat images that CEF never delivered under contention.