NVIDIA、JetPack 7.2.1 でエージェント型動画機能と T3000 エミュレーションを追加

Video is a core data path across NVIDIA Jetson applications, from robotics and intelligent video analytics to industrial automation, healthcare, media processing, and remote operations. A system may capture several cameras, decode network streams, run AI inference or conventional vision processing, draw results, and encode video for storage or delivery.

The individual calls are straightforward; the engineering work lies in choosing the right interface, codec, pixel format, memory path, rate control, and buffering for the Jetson device and proving that the complete path meets latency, throughput, and quality goals.

NVIDIA JetPack 7.1 introduced NVIDIA Video Codec SDK support on Jetson Thor, giving C and C++ developers direct, fine-grained access to NVIDIA Video Encoder (NVENC) and NVIDIA Video Decoder (NVDEC).

For the first time on Jetson, JetPack 7.2.1 adds support for PyNvVideoCodec 2.2, which is the NVIDIA Python library for hardware-accelerated video encoding and decoding on NVIDIA GPUs. PyNvVideoCodec produces and consumes video frames as GPU-resident device memory, exposed through the DLPack protocol and CUDA device buffers. It includes AI-pipeline-friendly features such as multi-mode frame sampling, ThreadedDecoder that strengthens pipeline efficiency by pre-decoding frames in background thread, decoupling decode latency from inference latency.

JetPack 7.2.1 also brings foundational agentic video skills above these SDKs. The SDKs supply the programmable video primitives; the skills connect a developer’s goal to live device discovery, supported configurations, working recipes, execution, measurement, and reproducible evidence. Together, they accelerate open models and agent development on Jetson, putting frontier-class AI performance in the hands of every developer, everywhere.

From prompt to verified pipeline

A prompt such as “How many H.264 1080p30 streams can this Jetson run for a low-latency use case?” can’t be answered from code generation or a data-sheet limit alone. It depends on the installed software, the target’s operational codec capabilities, the memory path, and a controlled run that proves latency and throughput.

Jetson video skills turn that intent into a repeatable codec workflow: Inspect the target, choose a supported Video Codec SDK or PyNvVideoCodec path, generate a tested configuration, execute it, and return measured results with warnings and evidence.

Jetson video skills provide an agentic workflow layer above Video Codec SDK and PyNvVideoCodec, which use the NVENC and NVDEC hardware engines:

Jetson Agentic Skills

Discover, Setup

Video Capability

Benchmark

Encoder Recipe

Pipelining

Others

Video Codec SDK

PyNvVideoCodec

NVENC

NVDEC

Foundational skills for Jetson video workflows

JetPack 7.2.1 introduces foundational video workflows through the unified jetson-videosdk skill. Developers can invoke each workflow independently or combine them to configure and verify codec stages within a broader pipeline. Future releases will expand this foundation with additional skills across Jetson developer workflows.

• Discover, set up, and report capabilities: Identify the Jetson platform and installed software, guide supported setup, and query encode and decode codecs, formats, memory paths, and session capabilities on the live target.

• Generate encoder recipes: Translate goals such as low latency, constant quality, constrained bitrate, resolution, frame rate, and codec choice into explicit settings and runnable Video Codec SDK or PyNvVideoCodec paths.

• Benchmark performance and quality: Run repeatable measurements for throughput, latency, utilization, bitrate, and output quality.

• Validate the codec workflow: Connect setup, recipe selection, encode or decode, measurement, and artifact handoff; then return the configuration, results, warnings, and evidence required to reproduce the outcome.

What the skills add above the SDK

The SDKs provide accelerated video primitives; the skills add device-aware configuration and verification. They distinguish advertised capabilities from operations that succeed on the target and preserve recipes, inputs, outputs, and measurements for reproducibility.

Build a AI video pipeline with PyNvVideoCodec and Jetson video skills

Consider the prompt, “Decode this video, apply my AI or computer-vision processing, and verify that the codec stage meets my performance target.” A coding assistant uses the Jetson video skill to configure and validate PyNvVideoCodec, while the sample applications provided with PyNvVideoCodec supplies a reference application scaffold.

Step 1: Inspect and configure. The coding assistant invokes the skill to verify the Jetson platform, software, and codec capabilities, then selects the decoder, output format, memory path, and buffering.

Step 2: Decode into framework data. A PyNvVideoCodec sample can prepare frames on a background thread and hand each frame to a framework tensor without application-specific codec-buffer code.

Step 3: Complete the application flow. With the codec stage configured, the coding assistant can extend the sample application with the stages required by the prompt, for example, preprocessing, detection or classification, privacy filtering such as blurring, visualization, and output handling. The exact AI or computer-vision model and application logic remain developer choices.

Step 4: Measure and verify. The skill records the codec configuration and verifies run status, throughput, latency, utilization, warnings, and evidence.

Scope note: The JetPack 7.2.1 video skills cover Video Codec SDK and PyNvVideoCodec. They do not add GStreamer, V4L2, AI-model, or application-level pipeline-building skills in this release. They complement a coding assistant at the codec stages. Additional Jetson skills in future releases can be composed with them to support broader end-to-end applications.

Choose the Jetson video interface for each pipeline stage

Jetson exposes video through complementary software layers. GStreamer provides high-level pipeline composition, V4L2 provides Linux video device and buffer controls, Video Codec SDK provides lower-level C/C++ access to NVENC and NVDEC, and PyNvVideoCodec builds on the core Video Codec SDK APIs with simpler Python interfaces. These layers can be combined when a workflow crosses capture, codec, AI, and delivery stages.

*Table 1. Match the interface to the pipeline layer requiring the most control; interfaces can be combined within a single application*

Together, JetPack 7.2.1 gives developers a clearer path from intent to evidence: choose the appropriate Linux, C/C++, or Python interface; keep accelerated data close to the GPU; and use foundational video skills to configure and verify the Video Codec SDK or PyNvVideoCodec stages. This foundation complements coding assistants today and can combine with additional Jetson skills in future releases to support broader end-to-end applications.

Emulate Jetson T3000 performance on a Jetson T5000 with JetPack 7.2.1

The Jetson T3000 delivers 865 FP4 TFLOPS in a compact, power-efficient platform for humanoid and robotics workloads. It provides inference performance comparable to T5000 for multimodal AI while reducing footprint, power consumption, and cost. With JetPack 7.2.1 now you can kick start your development by emulating the recently announced T3000 performance over a Jetson T5000 module of the Jetson Thor AGX Developer Kit. Visit the Jetson Linux Developer Guide for more details on the implementation of the emulation.

Get started

• Jetson device skills repository

• Jetson BSP skills repository

• Video Codec SDK developer portal

• PyNvVideoCodec Get Started

• PyNvVideoCodec Programming Guide

• Jetson Linux Developer Guide for T3000 emulation

• Jensen Huang on Open Models and American AI Leadership

• Jensen Huang on the Open Secure AI Alliance

• Jetson T3000 and T2000 launch blog

• Jetson Linux Multimedia API reference