Current workload limits and evidence
Separate long historical capacity runs, profile admission limits and the latest short moving-browser transition checks.
Reviewed 15 October 2026 against e72ff8b4. Setup defaults and capacity notes updated at 74e3392f. These describe code and named workload settings, not the live state of every deployment.
The fixed-pool export Electron and the mixer's copy on import were removed at 159d0182: browsers run stock Electron with Chromium's ordinary capture, sampled in place. Measurements below that name a fixed pool or a copy on import describe the removed path.
What Setup selects now
The nvidia_l4_cuarray profile sets the browser-window allowance by canvas rate: 40 at 25, 30 and 50 fps; 30 at 60 fps. These totals include four DSKs, leaving 36 or 26 full-size scene-browser pages. Both the Setup interface and backend apply the rate-specific limit; changing the rate does not silently retain the older forty-window allowance at 60 fps.
AUX encoders default to p1. Explicit saved encoding choices remain editable and are preserved; choosing Defaults restores p1. This is a default for newly generated configuration, not a forced migration of existing shows. Browser workers accept up to ten windows each, and the L4 deployment keeps 32 compute/eight copy queues.
The fresh L4 seed remains 60 fps HDR 4:2:2, now with 78 sources: 44 decoded video inputs, four v210 inputs, 26 scene-browser pages and four DSKs, with 17 outputs and AUX p1. That seed is a setup configuration, not a new measured capacity result. It differs from the 4:2:0, 57-video, 18-output transition workload below.
Browsers run stock Electron with Chromium's ordinary capture, and the mixer samples each captured buffer in place. The fixed nine-buffer exporter and the copy per browser frame of the transition checks below have been removed.
Moving browser workload · 15 October 2026
12-vCPU NVIDIA L4; 1080p HLG 4:2:0; 57 video inputs; 18 encoded outputs. Counts include four cropped DSKs; every other browser page is animated at 1920×1080 and the canvas rate. Fixed prewarmed pool nine, one CUDA copy per browser frame, 32 compute/eight copy connections and L4 frame staggering enabled. Each row covers four transitions between two 32-layer grids: one-second fades/cuts six seconds apart.
| fps | Windows / workers | AUX preset | Browser repeats / skips | Key repeats / skips |
|---|---|---|---|---|
| 25 | 40 / 4 × 10 | p3 | 0 / 0 | 0 / 0 |
| 30 | 40 / 4 × 10 | p3 | 0 / 1 | 0 / 0 |
| 50 | 40 / 4 × 10 | p3 | 0–3 / 0–3 | 0 / 0 |
| 60 | 30 / 3 × 10 | p1 | 3 / 3 | 0 / 0 |
The table preserves the presets actually measured: AUX p3 at 25/30/50 fps and p1 at 60 fps. The current Setup default is p1 at all those rates; changing a default does not rewrite the historical result. Main presets: SDR p5, HDR p3, clean p3. AUX p1 is a quality/throughput tradeoff. These runs had no compositor misses or new imports. A second recorded four-fade pass at 50 fps had zero observed browser/key repeats or drops; at 60 fps it had one browser repeat/drop and two skipped key frames. Encoded recordings had regular RTP cadence and changing sampled ticker, corner and grid regions.
These are short smoke results, not zero-drop guarantees. They do not qualify 64-layer transitions. At 60 fps, 32 windows repeated hundreds of frames and 36/40 collapsed into persistent repeats. Three larger worker groups failed the 40-window/50-fps case; grouping is part of the result.
Historical capacity and admission
The earlier 16-vCPU L4 192-source SDR30 show used 105 NVDEC, 40 mostly still browser and 47 raw inputs, with 24 outputs. It ran 40 minutes with no Program/AUX deadline miss or AUX drop. That workload is different from the moving-page transition checks above. Other format/rate rows in the profile are admission limits unless explicitly measured.
The maintained capacity reference retains those distinctions and GPU-specific tables. Do not extrapolate L4 numbers to T4, or treat a successful code/pixel test as a new capacity benchmark.
Configuration is not deployment state
The reviewed branch includes these implementations and L4 defaults. The Docker image and direct npm installs run stock Electron plus shim. This page does not assert that a public demo or a different host is running that build. Check its actual images, recipe, environment and reported pool before comparing results.