Gaming performance: measurements and next experiments¶

Reviewed 18 October 2026. Change one thing, replay the same workload, and compare results beyond the run-to-run spread. Current per-emulator decisions and measurements live in emulator notes; earlier standalone Flycast investigations are in the bug history.

Current baseline¶

Games already default to cores 3–7 in game.slice; system services and unbound workqueues use 0–2 during gameplay. PS2 pins EE to the X3 and VU/GS to the A715s, with its frontend on 5–6. Verify with coremap-check; see CPU isolation.
The gaming GPU floor is already 401 MHz (gpuperf), restored on exit. The earlier proposal to raise it from the idle floor has been implemented.
RetroArch, the launcher and Steam use integer-multiple VRR within 65–120.198 Hz. Steam is capped at 60 fps; xemu and mpv use fixed refresh. No vsync or BFI setting is available. See variable refresh.
Flycast and Dolphin run as libretro cores. Standalone backend settings, start_flycast.sh, and sway experiments no longer apply.
RetroArch defaults to an 8 ms audio request with a 3 ms quantum floor. Shared-memory playback, native link rates and graph tuning are already in the image. See audio latency.
NeoCD tuning is finished. Its measurements expose remaining host policy costs, especially real-time threads holding high clocks on lightly used cores.

Record a comparable run¶

Use a fixed scene or a validated gamecapture workload. Capture verification must match the core build used to record it; a mismatch is not evidence that an optimization broke determinism. Keep ROM, BIOS, core, shader, scaling and input sequence fixed, and retain intro/loading and gameplay as separate windows.

Enable MangoHud before launching, on the Nova:

. /etc/profile
set_setting portareos.mangohud.enabled 1

To record frame intervals, configure MangoHud logging as described in the debug tools guide. Showing the overlay alone does not start a log. With the game and its recording running:

perf-probe baseline-01 240
cat /storage/perf/baseline-01/summary.txt

Use a new label for each run. perf-probe writes /storage/perf/<label>/:

File	What it establishes
`env.txt`	Kernel command line, CPU/GPU bounds and governors, forced PipeWire metadata and any legacy emulator configs found
`samples.csv`	Once-per-second CPU clocks, including middle and prime clusters, GPU clock, temperature, battery charge, emulator CPU ticks and xrun counters
`frametimes.csv`	Newest non-summary MangoHud CSV, when available; confirm it belongs to this run
`summary.txt`	Available clock, thermal, charge, xrun and frame-interval summaries; gawk is needed for full percentile processing

The probe does not save every current RetroArch setting or start MangoHud logging. Save the launched RetroArch config, core options, relevant system.cfg overrides, image/core identifiers and capture metadata alongside the run. Forced PipeWire metadata is not a substitute for the running sink's rate and quantum: record those with pw-top.

Run each candidate and its baseline at least three times. Compare application_interval_ms (between presents) with display_interval_ms (between frames reaching the panel); blank display timing is missing data. Keep p95/p99, late/repeated-frame counts, xruns and CPU cost. A mean FPS alone hides hitches. Watch temperature and clock drift during longer runs, and measure drain on battery; the gauge cannot establish drain while charging.

For display scheduling use display-trace and display-check. For pad poll to first scanline use the latency trace markers; see latency per system. Neither is physical button-to-screen latency.

Next experiments¶

Real-time threads and clock floors¶

NeoCD's main emulation thread avoids the X3, yet the prime core stays at 2957 MHz for mostly idle real-time audio/GPU threads. The default RT clamp is 1024. Three measured shortcuts failed: setting the default to zero made late frames worse; restricting RetroArch to 3–6 or moving the GPU worker to 0–2 increased CPU cost and left the X3 high. Do not repeat them as untested fixes; NeoCD holds the numbers.

Test per-thread clamping or a kernel policy that lets lightweight RT work scale without slowing the emulation threads. Compare clocks, CPU cost, xruns, read-to-first-line latency and delivered frames on both light and heavy cores. Keep PS2's performance governor until schedutil with its thread clamps matches the recorded performance.

GPU floor and shader cost¶

Compare the 401 MHz gaming floor with lower available frequencies per system, keeping the renderer and shader fixed. Do not raise the ceiling as a first response to a lightly loaded GPU. Save original values and restore them after each run; check the launcher's idle floor after exit.

A shader's GPU work is latency on the timed-present path. Compare presets with both delivery and latency traces, not utilization alone. The SNES CRT shader and handheld LCD shaders already have measured costs in LATENCY.md; NeoCD's scanline shader adds about 0.6 ms in its recorded test. Preserve the chosen geometry and appearance when evaluating an alternative.

Flycast regression and Steam validation¶

Replay Tony Hawk's Pro Skater under the current Flycast core before carrying the standalone stutter forward. If it reproduces, trace the current RetroArch audio and presentation paths, without editing removed backend configs. Its intended 30 fps is not itself a slowdown.

Run a real Steam game on the final kernel/gamescope combination. Verify the 60 fps ceiling through focus changes and Steam's frame-rate overrides, check repeat cadence and late frames, and return to the launcher. The existing probe tests do not establish that game-level result.

Acceptance¶

Keep a change only when the same validated workload shows an improvement larger than run-to-run spread while audio latency, xruns, frame delivery, controls and suspend/resume remain acceptable. Record rejected experiments as well as successful ones. Restore experimental settings after measuring; shared defaults need evidence from every affected class of emulator.