Jetson Video Benchmark
NVIDIAOfficialCreated by Vinit BansalDistributed by NVIDIA
- Skill
- Workflow
- Software development
- Apache-2.0
- Medium risk
AutoGineer summary
Written by AutoGineer from the skill’s own files, not by the author
The jetson-video-benchmark skill provides a structured workflow for measuring the encode and decode throughput of the Jetson Video Codec SDK and PyNvVideoCodec. It supports live performance measurements using authenticated samples, preset comparisons, and codec-worker capacity sweeps. For scenarios where representative media is unavailable, the skill generates clearly labeled planning estimates based on official documentation. It enforces strict validation of inputs, environment readiness, and output reporting to ensure results are reproducible and verifiable.
What it does
- Measures encode and decode FPS and megapixels per second for specific workloads.
- Compares performance across different presets and surfaces.
- Tests codec-worker capacity through systematic sweeps.
- Generates performance estimates based on documented SDK references when media is absent.
- Validates benchmark results against strict output contracts and sample markers.
- Reports detailed metrics including mean, minimum, and maximum performance values.
When to use it
- When you need to measure actual encode or decode throughput on a Jetson device.
- When comparing the performance impact of different codec presets or surfaces.
- When determining the capacity of codec workers for a specific workload.
- When you require a performance estimate based on official documentation for planning purposes.
- When you need to verify performance metrics using authenticated SDK samples.
Advantages
- Ensures consistent and verifiable benchmark results through strict output contracts.
- Provides clear separation between live measurements and documentation-based estimates.
- Automates the preflight checks for environment and dependency readiness.
- Supports both native C++ SDK samples and Python-based PyNvVideoCodec workflows.
- Prevents invalid comparisons by enforcing strict input and configuration validation.
Usage examples
- “Measure the decode throughput for this H.264 elementary stream file: /data/input.h264”
- The agent validates the input file, checks the environment, runs the AppDecPerf sample, and reports the measured FPS and MP/s.
- “What is the expected encode performance for 1080p60 H.265 on Jetson Orin?”
- The agent provides a documented performance estimate based on the SDK reference, clearly labeling it as a non-measured planning value.
- “Compare P4 and P5 preset performance for this raw video file: /data/raw_frames.yuv”
- The agent runs the encode benchmark for both presets using the provided raw input and reports the throughput difference.
Install
Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/jetson-video-benchmark/downloadUnzip it into the cross-client skills folder of your project:
mkdir -p .agents/skills && unzip jetson-video-benchmark.zip -d .agents/skills
Documentation by Vinit Bansal
Jetson Video Benchmark
Purpose
Measure codec-stage FPS and megapixels/second for an evidenced target and exact workload. Supported live routes are encode, decode, P4/P5 comparison, and a strictly increasing worker-capacity sweep. A separate no-media path calculates clearly labeled SDK-documentation estimates; it never claims a measurement.
Terminal gates
For a request solely for PSNR or SSIM, state that objective quality is outside this skill and needs a separately authorized workflow, then stop. Return only that scope response. Do not append an alternative benchmark or next step, name another tool, request media, probe, or install anything.
Resolve codec direction before probing or calculating. Explicit encode or decode wording wins. Otherwise, quality, preset, bitrate, rate control, recording, or compressed-output wording implies encode; explicitly compressed input, ingest, playback, or IP/RTSP input implies decode. Treat generic camera-count wording such as "connect" or "handle", and a codec name on a camera or device, as direction-neutral. Set neutral wording aside and resolve direction on the remaining cues; a neutral cue never creates a conflict. For an encode-led camera estimate, state that NVDEC applies only when cameras already emit the named compressed codec, keep encode primary, and ask the user to confirm direction after giving the planning scenarios. If non-neutral cues conflict or are absent, present both interpretations, state that a unique encode estimate also needs an exact preset and measured validation needs representative input, then ask which direction applies and stop at that gate. Transcoding consumes separate decode and encode budgets: measure it as the supported decode and encode routes and report each budget separately, never as one combined FPS.
An explicit live/run/real/actual measurement requires representative input for every requested direction: raw frames with format, geometry, and frame count for encode; a compressed elementary stream or container for decode. Each input must be one exact target-local path or user-supplied HTTP(S) URL. If any is absent, return only
input_requiredand ask for the missing item(s), including a separate exact path or URL for every requested direction, before reading workflow references or sibling skills, probing, authentication, recipe work, or workspace creation. Never benchmark the setup smoke fixture or substitute catalog media.A no-media planning/expected/indicative question follows documented-performance-estimates.md. It requires the exact documented row and, for a scaled estimate, clock provenance; it sets
measurement_performed: false. Preserve that reference byte-for-byte. Omitted preset, per-stream FPS, or stream mix does not block planning: enumerate the reference's bounded documented candidates and scenarios, disclose every assumption, and ask for the omitted values.
Live preflight
Read benchmark-workflow.md completely once,
then apply its preflight section. Preserve explicit
native, pynvc, and both; map delegated selection to auto. Obtain fresh
read-only readiness from jetson-video-setup, passing any exact interpreter
already supplied or established in this conversation. Setup otherwise checks
the conventional profile path. For auto, zero eligible surfaces blocks, one
runs, and two returns selection_required. Python encode and comparison need a
full-samples environment; decode performance may use pynvc-smoke.
Encode, compare, and encode-capacity also require a validated schema-2 recipe
from jetson-video-recipe. A missing dependency stops that branch with
dependency_required; an eligible peer may continue as partial.
Measure
Follow that same reference for sample allowlists, builds, direct argument lists, frame accounting, option checks, marker grammar, comparisons, and worker sweeps.
Show the dry-run plan, then use a fresh mode-0700 workspace. Authenticate each launcher immediately before use and launch its literal argument list.
Run one excluded whole-process warmup and at least three new measured processes per variant and surface, each with a finite timeout and without
-loop.Accept the sample-reported FPS only when its positive marker and processed frame count match. Apply every identity, workload, marker, and statistics check in benchmark-output-contract.md.
Report
Retain each launch's unedited output and write the compact result described by benchmark-output-contract.md. Report every repetition, recomputed mean/minimum/maximum, MP/s only when dimensions are sample-bound, partial branches, limitations, and retry reason. Label worker sweeps as codec-stage capacity bounds and preset comparisons as throughput-only.
Limitations
Results apply only to the evidenced target, release, clocks, sample, content, frame range, and controls; they are not a portable product ceiling.
Do not interpolate undocumented presets or scale across format, bit depth, chroma, codec, rate control, or tuning. Resolution scaling is allowed only as the estimate reference's explicitly labeled pixel-area heuristic.
Keep documented estimates per engine; use an all-engine multiplier only under the explicit aggregate-planning rule in the estimate reference.
The PyNvVideoCodec 2.1 encode-performance helper caps each worker at 1,000 frames; follow the common-frame rule in the workflow.
Preserve
input_required,selection_required,dependency_required,blocked,partial, andfailed; do not promote a peer's success.
Files
- ATTRIBUTION.mdAdded by AutoGineer930 B
- BENCHMARK.md7.8 KiB
- LICENSELicence file added by AutoGineer11.2 KiB
- SKILL.md6.4 KiB
- evals/evals.json6.2 KiB
- references/benchmark-output-contract.md3.6 KiB
- references/benchmark-workflow.md10.1 KiB
- references/documented-performance-estimates.md12.3 KiB
- skill-card.md4.9 KiB
- skill.oms.sig5.2 KiB
evals/evals.json
{
"skill_name": "jetson-video-benchmark",
"evals": [
{
"id": "benchmark-no-content",
"prompt": "Can you measure real 1080p H.264 encode and decode FPS without a test video?",
"expected_output": "A request for the missing representative raw encode input and compressed decode input before any real benchmark.",
"assertions": [
"Says real performance cannot be measured without representative input",
"Asks for exact target-local paths or user-supplied HTTP(S) URLs for raw encode input and compressed decode input",
"Does not choose content or claim measured FPS"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
},
{
"id": "benchmark-documented-estimate",
"prompt": "Estimate H.264 1080p P1 low-latency encode FPS and H.264 decode FPS on this Jetson without running a benchmark.",
"expected_output": "A documentation estimate scaled with the target's configured maximum video clock and clearly labeled as unmeasured.",
"assertions": [
"Discovers the target's configured maximum video clock instead of requiring it from the prompt",
"Uses the SDK 13.0 reference rows and the detected-clock-to-1691-MHz scaling formula, treating a clock reported within 1 MHz of 1691 MHz as the same reference operating point",
"Labels the result as an estimate and does not claim measurement or measured or verified concurrent-stream capacity"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
},
{
"id": "benchmark-undocumented-p4",
"prompt": "What is the documented H.264 1080p P4 low-latency encode FPS on Jetson Thor?",
"expected_output": "Documentation estimate unavailable because the SDK 13.0 Thor table has no P4 row.",
"assertions": [
"States that the SDK 13.0 Thor table has no exact P4 row",
"Does not interpolate between other presets or call P4 unsupported",
"Does not claim measured FPS"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
},
{
"id": "benchmark-4k-planning-estimate",
"prompt": "Without running a benchmark, estimate HEVC P3 VBR high-quality 3840x2160 encode FPS on Jetson Thor at a configured 1557 MHz maximum video clock, then give a 30-fps per-engine stream planning bound with 10% headroom.",
"expected_output": "A clearly labeled theoretical estimate of about 133.28 FPS and a conservative three-stream per-engine planning bound, followed by a request for representative content to measure it.",
"assertions": [
"Uses the exact SDK 13.0 HEVC P3 VBR high-quality 1080p row, 1557/1691 clock ratio, and 1080p-to-4K pixel ratio",
"Labels inverse-pixel scaling as a heuristic not documented resolution performance, with measurement_performed false and capacity_verified false",
"Applies the 0.9 margin without engine multiplication, does not claim a verified camera count, and asks for representative content for a real benchmark"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
},
{
"id": "benchmark-natural-camera-capacity-planning",
"prompt": "How many 4K cameras and 1080p H.265 cameras can I connect to this Jetson for high video quality?",
"expected_output": "Theoretical codec-stage planning scenarios built from the documented HEVC VBR high-quality rows and pixel-area scaling, with no single invented stream count while per-stream FPS and mix are unspecified, then a request for FPS, mix, and representative content.",
"assertions": [
"Enumerates the documented HEVC VBR high-quality preset candidates instead of silently choosing one, and does not return documentation_estimate_unavailable merely because preset, FPS, or mix was omitted",
"Reads 4K as a disclosed 3840x2160 planning interpretation and shares one budget across resolutions instead of granting each resolution the full engine",
"Claims no verified camera count, keeps measurement_performed and capacity_verified false, and asks for per-stream FPS, stream mix, and representative content"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
},
{
"id": "benchmark-camera-direction-medium-quality",
"prompt": "How many 1080p @ 60 fps h264 camera can i connect to this Jetson for medium quality?",
"expected_output": "An encode-led theoretical camera-capacity answer because quality is an encode cue, with decode kept as a conditional interpretation only for cameras that already emit H.264, no silent mapping of medium quality to one preset, and a request for the exact direction, preset, and representative content.",
"assertions": [
"Leads with NVENC encode planning and states that NVDEC applies only if the cameras already emit compressed H.264; it does not present decode capacity as the primary answer",
"States that medium quality is not an exact documented preset or tuning, does not silently select one, and keeps any documented encode rows as clearly labeled candidates",
"Claims no verified physical-camera count, keeps measurement_performed and capacity_verified false, and asks for codec direction, exact preset, and representative content"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
},
{
"id": "benchmark-quality-boundary",
"prompt": "Use the jetson-video-benchmark skill to answer this. Give me the PSNR and SSIM quality numbers for H.264 vs AV1 encode.",
"expected_output": "A strict performance-skill scope boundary with no quality numbers, external-tool recommendation, or operation.",
"assertions": [
"States only that this performance skill does not provide PSNR or SSIM and that a separately authorized quality workflow is required",
"Does not fabricate or defer quality numbers and does not ask for media, probe, encode, or install anything",
"Does not name or recommend FFmpeg or another external quality tool and does not offer to configure or run one"
],
"expected_skill": "jetson-video-benchmark",
"expected_workflow": "benchmark"
}
]
}
Requirements and compatibility
Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta
“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.
Security
What our scanner noticed· Medium risk
- Large base64-encoded blobs
A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.
Attribution and licence
- Created by
- Vinit Bansal
- Publisher
- NVIDIA (Official)
- Distributed by
- NVIDIA
- Source
- NVIDIA/skills · skills/jetson-video-benchmark · commit fce3cca
- Licence
- Apache License 2.0 (redistribution allowed)
- LICENSE (added by AutoGineer from LICENSE-APACHE)
- ATTRIBUTION.md (written by AutoGineer)
- Synced
AutoGineer is not affiliated with or endorsed by Vinit Bansal or NVIDIA.
Download ZIP
The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download jetson-video-benchmark.zip