Follow-up with better-controlled data. Two corrections to my first post: I hadn’t equalised render resolution between the two drivers, and I hadn’t checked Proximity Sensor Type.
Setup: Pimax Dream Air (SLAM), SBoys3 1.3.0, SteamVR 2.17.10, Elite Dangerous (OpenVR). RTX 4080 Super on driver 32.0.16.1692, Ryzen 7 9800X3D, 72Hz for a 13.89ms frame budget. Fixture is the same every time: parked on an outpost landing pad, roughly 60 second fpsVR capture.
Native Pimax driver (aapvr) at full resolution gave a GPU median of 13.00ms with p95 at 13.90ms, and reprojected 0 of 19,881 frames — zero percent, zero dropped.
SBoys3 with render multipliers at 0.95 gave a GPU median of 12.90ms with p95 at 13.30ms, and reprojected 3,376 of 14,820 frames — 22.8 percent, 4 dropped.
SBoys3 with render multipliers at 0.90 gave a GPU median of 11.70ms with p95 at 12.00ms, and reprojected 886 of 15,758 frames — 5.6 percent, 2 dropped.
The finding: SBoys3 renders faster than the native driver at every frametime percentile — lower median, lower p95, lower max — while rendering fewer pixels, and still reprojects where the native driver reprojects nothing. Dropped frames sit near zero throughout, so the compositor appears to be predicting misses that never occur rather than reacting to real overruns. Put another way, native reaches zero percent reprojection at full resolution and 13.00ms, while SBoys3 needs roughly 19 percent fewer pixels to get down to 5.6 percent and still doesn’t reach zero.
It’s reproducible on toggle. Switching SBoys3 off and back on moves reprojection between 0 percent and roughly 23 percent with no change in GPU frametime. I tried both the Dream Air Custom Default and the Dream Air Like Pimax distortion profiles and saw no difference.
Config, in case any of it matters: Parallel Projection is on because IPD geometry breaks without it. Hidden Area Meshes on, Stationary Dimming on, Proximity Sensor Type set to Stationary Dimming, FOV Clamping on, Rendered IPD 65.7, Super Sampling Filter Percent 500, and the VSync-to-Photons and Photons-to-Vblank values left at stock 0.007 and 0.0025. On the SteamVR side, Motion Smoothing is disabled, Advanced Supersample Filtering is disabled, and resolution is at 100 percent. Eye tracking is off in Pimax EVO but on in the SBoys GUI, and I’m not sure which one actually governs.
Three questions.
First, is 5 to 25 percent reprojection at these frametimes expected behaviour on 1.3.0, or does this look like a frame-pacing problem?
Second, should the eye tracking toggle in EVO and the one in the SBoys GUI agree with each other?
Third, are the stock VSync-to-Photons and Photons-to-Vblank values correct for the Dream Air, or do they want tuning per headset?
Happy to run anything else against the same fixture. I have raw fpsVR CSVs and session JSONs for every run.