Imported from nvidia/skills (skills/jetson-video-benchmark/SKILL.md). Install upstream with npx skills add nvidia/skills --skill jetson-video-benchmark. Copyright stays with the author (Apache-2.0).

Jetson Video Benchmark

Purpose

Measure codec-stage FPS and megapixels/second on the current Jetson. Use this skill for encode or decode throughput, P4/P5 comparisons, native-versus-Python comparisons, and increasing-worker capacity tests. When content is unavailable, it can instead produce a clearly labeled SDK-documentation estimate for an exact supported 1080p table row and target maximum video clock. For another requested resolution, it may additionally apply the bounded pixel-area heuristic defined in the estimate reference. Never present either estimate as a target measurement.

Prerequisites

• For a live measurement:
  • Run on the target Jetson with direct GPU access. A fresh validated nvcodec-environment identity from jetson-video-setup is optional; when supplied, it is authoritative and any invalid or stale identity fails without local fallback. The agent may obtain that identity from setup's public read-only probe; it need not be present in the customer's prompt.
  • Without that identity, native routes inspect only the installed nvidia-video-codec-sdk APT package and its package-owned official sample sources. PyNvVideoCodec routes require an authenticated setup environment or the caller's exact absolute pynvc_interpreter; never scan for a venv. Before asking the customer for that path, invoke setup's public probe when that skill is installed and inspect its typed result. These read-only checks install, repair, register, and smoke-test nothing. Only when neither authority is usable does an explicit pynvc or both request return input_required; local auto records PyNvVideoCodec as not_evaluated and may continue an eligible native branch.
  • A PyNvVideoCodec decode-only performance route can use a validated default pynvc-smoke environment. A Python encode or compare route uses official samples that import Torch and therefore requires a separately provisioned full-samples venv; never upgrade the smoke venv in place.
  • Live encode, compare, and encode-capacity routes require sibling jetson-video-recipe and an exact portable identity for one of its validated schema-2 recipes. If its public validator is absent, preserve dependency_required and its install-and-retry action. Decode routes do not require the recipe skill.
  • When setup is installed, read its shared video content policy before a live measurement. It does not apply to the separate documentation-only estimate path. Setup is not required solely for this policy: without it, require one exact user-selected path or URL, never choose catalog or synthetic media, and preserve source URL, license, attribution, path, size, and SHA-256.
  • Apply that policy's input gate before constructing a live benchmark dry run. Never choose media for the user or use the setup smoke fixture for performance.
• A documentation answer may be produced off-target when the exact platform, SDK version, table conditions, and configured maximum video-clock facts are supplied with provenance. A clock-scaled estimate also requires a positive configured maximum video clock. It requires no media, recipe, sample authentication, or codec launch. Without that clock, report only the unscaled reference row as target_clock_unavailable, not a platform-scaled estimate.

Compose requested sibling stages

Documentation-only estimates and recipe-free live decode do not require a sibling skill. Live encode, compare, and encode-capacity routes require jetson-video-recipe; SDK installation, repair, a new full-samples Python environment, or one read-only handoff when registered Python authority is required belongs to jetson-video-setup. Check the agent's installed skill catalog before either stage. If the sibling is present, read its SKILL.md and invoke its documented public entry point; pass artifacts as data and never import sibling code. If it is absent, preserve completed input and measurement evidence and say, using the actual names: I can run <stage>, but it requires <skill>, which is not installed. Install <skill> and retry this stage. Never require setup when the caller already supplied authenticated setup evidence or an exact interpreter that passes local authentication, or a recipe for a decode-only or documentation-only request.

Instructions

1. Apply the scope boundary first. For a request solely for objective quality metrics, including PSNR or SSIM, state only that this performance skill does not provide them and that a separately authorized quality workflow is required, then stop. Do not name or recommend an external tool, and do not offer to configure or run the comparison; do not request media, probe, install anything, or launch an operation.

2. Classify an in-scope request as a live encode, decode, compare, or camera_capacity measurement, or as a documented_estimate answer. Resolve camera data direction before selecting an encode or decode row. Quality, preset, bitrate, rate-control, recording, or requested codec-output wording is an encode cue; lead with NVENC and mention decode only as the conditional case where cameras already emit the named compressed codec. IP/RTSP input, already-encoded input, ingest, playback, or explicit decode wording is a decode cue. Explicit transcoding or decode-then-encode uses separate NVDEC and NVENC budgets. When no direction cue exists, present the encode and decode interpretations conditionally and ask which applies; never silently choose one.

3. When media is absent:

   • For an explicit request to run or benchmark, or for actual, real, measured, live, or on-this-target FPS, apply the shared policy's input_required gate and stop before probing, retrieval, authentication, workspace creation, dry run, or launch. Ask only for the missing media identity at this gate; do not also request an interpreter, environment, recipe, or other later-stage field.
   • For a planning, expected, indicative, or achievable-FPS question at any positive requested resolution, follow Documented performance estimates. Use only an exact documented 1080p row and a configured maximum video clock as the source basis. For a non-1080p request, apply the reference's inverse-pixel-area formula and disclose that it is an additional heuristic not stated by the SDK table. If the clock is unavailable, return only the unscaled 1080p reference row. Do not invoke the benchmark controller or claim a measurement.

4. When local media is supplied, validate and hash that exact selected input. For URL input, preserve the exact user-supplied URL and retrieve it only after the target, authorization, and runtime-authority gates; then validate and hash the retrieved bytes. Preserve source URL (or null for local media), license, and attribution. Use the live measurement path, not a documentation estimate.

5. Preserve the requested surface for live measurements. Treat “whichever”, “best available”, “choose for me”, and other unspecified-surface wording as auto, never as both. Reserve both for an explicit dual-surface comparison. For auto, zero eligible surfaces block, one runs, and two return selection_required; never inspect prior results, invent a preferred SDK, or launch a benchmark to make the missing user choice. With two eligible surfaces, ask for exactly native, pynvc, or both. After the media gate, use valid supplied setup evidence or an exact interpreter first. If PyNvVideoCodec may participate and neither is supplied, invoke installed jetson-video-setup through its public read-only probe_nvcodec.py: use --runtime pynvc for explicit Python or --runtime both for both/auto, a fresh --output, and never --setup-candidate. Inspect the fresh artifact; only a live artifact whose selected Py surface is installed and whose pynvc.identity.status is verified is usable authority. Snapshot that exact artifact as the controller's portable environment identity with exactly schema_version, kind, canonical absolute path, size_bytes, and lowercase sha256. If setup is absent or reports any not-ready, unreadable, stale, binding, or launch failure, request the exact interpreter for explicit pynvc/both; for local auto, report Python as not_evaluated and continue only an eligible native surface. Never pass a blocked probe as authority, scan for a venv, or hide an unevaluated peer.

6. Run a dry_run first, review every planned argument, then run execute in a fresh private workspace. Invoke this skill's controller directly:

   python3 -I {baseDir}/scripts/benchmark_controller.py \
     --request request.json --workspace fresh-workspace \
     --output result.json

7. For native encode, authenticate AppEncPerf, inspect both -h and -A, and use only advertised options. Never pass -loop. Native decode uses authenticated AppDecPerf; PyNvVideoCodec uses the wheel-owned performance samples under either the setup-evidenced interpreter or the exact absolute pynvc_interpreter selected by the caller.

8. Retain one whole-process warmup and at least three separate measured processes per variant and surface. Every measured command phase is measure.

9. For a live measurement, list every repetition's FPS and MP/s, then the mean, minimum, and maximum of both metrics. PyNvVideoCodec decode omits MP/s because its authenticated performance sample does not report dimensions; report that reason instead of deriving MP/s from caller metadata. For a native/Python comparison, disclose exact projection differences and whether they changed the requested intent. For a live preset comparison, report only measured throughput differences; do not state or imply a quality, compression-efficiency, or storage ordering from preset names or throughput. State that quality was not measured when that distinction matters. For a documented estimate, report the source row, clock provenance, requested resolution, formula, assumptions, and every inference label required by the estimate reference, plus measurement_performed: false; never invent repetitions or measured statistics.

10. Label measured concurrency results as codec-stage capacity bounds. For a no-media planning question, a resolution-scaled estimate may additionally produce a documented_theoretical_capacity_estimate when per-stream FPS and an explicit or clearly defaulted safety margin are available. Keep encode and decode budgets separate; for mixed workloads, sum each stream's fractional load against one shared budget instead of granting every resolution the complete budget. Call the result a theoretical codec-stream bound, never a verified camera count: it excludes capture, ISP, transport, AI, display, memory contention, and end-to-end latency. Disclosed 30-fps and 60-fps scenarios may stand in for an unstated frame rate, but they do not replace the question: the final response must still explicitly ask for every omitted input it assumed or enumerated, naming the codec direction (encode captured frames, decode already-compressed streams, or both), exact preset, per-stream FPS, and stream mix or count. End by asking for exact representative content and a measured, strictly increasing worker sweep to verify capacity. Use a read-only calculator for every documented FPS and stream-count calculation, preserve full precision until the final floor or display rounding, and apply the reference's two-sided maximum-count check; never rely on mental arithmetic for a reported capacity.

References

• Benchmark workflow defines routes, preparation, and execution.
• Benchmark output contract defines result states, repetitions, frame accounting, and statistics.
• Documented performance estimates defines the no-media SDK 13.0 table, clock and pixel-area scaling, theoretical stream budgeting, and reporting contract.

Available Scripts

Inspect the controller's public CLI directly:

python3 -I {baseDir}/scripts/benchmark_controller.py --help

Limitations

• Results apply only to the evidenced target and workload; they are not a portable product ceiling.
• Documentation-derived and resolution-scaled values are indicative per-engine planning estimates, not achieved FPS, verified concurrency capacity, or a substitute for testing representative content.
• Do not interpolate undocumented presets or scale across formats, bit depths, chroma, codecs, rate-control modes, or tuning. Scale resolution only through the explicitly labeled pixel-area heuristic. Do not multiply by an engine count unless that exact target count is independently authenticated, the user explicitly requests multi-session aggregate planning, and the answer remains a separately labeled multi-session theoretical bound.
• Objective quality measurement, including PSNR and SSIM, is outside this performance skill.
• The released PyNvVideoCodec 2.1 encode performance helper caps each worker at 1,000 frames. Follow the output contract rather than silently comparing unequal frame counts.
• External codecs and wrapper wall time cannot substitute for official sample-reported throughput.

Troubleshooting

• Preserve input_required, selection_required, blocked, partial, and failed instead of promoting them to completion.
• Reject stale outputs, input or environment identity drift, malformed or non-finite metrics, wrong processed-frame counts, unsupported sample options, and missing positive markers.
• Retain a successful surface when its peer fails, but never pool or substitute their results.