jetson-knowledge

name

jetson-knowledge

description

NVIDIA Jetson single-platform reference — module specs and AI performance (Orin Nano / Orin NX / AGX Orin / Xavier NX / legacy Nano), the JetPack / Jetson Linux (L4T) software stack, Super Mode power profiles, and the TensorRT engine-build workflow (.onnx → .engine via trtexec / nvpmodel / jetson_clocks). Use whenever someone asks about Jetson hardware specs, "多少 TOPS", JetPack/L4T versions, how to flash or set up a Jetson, or how to run a model on Jetson with TensorRT. 触发词:Jetson、Orin Nano、Orin NX、AGX Orin、Xavier NX、Jetson Nano、多少 TOPS、Super 模式、JetPack 版本、L4T、刷机、tegrastats、jtop、TensorRT、trtexec、engine 转换、nvpmodel、jetson_clocks、CUDA 版本。Routing — comparing Jetson against RDK / Raspberry Pi / RK3588 or "买 Jetson 还是 RDK" selection → rdk-ecosystem (cross-platform comparison lives there, anchored on RDK); Raspberry Pi specifics → rpi-knowledge; Rockchip RK3588 specifics → rk-knowledge. This skill covers Jetson itself only.

NVIDIA Jetson Knowledge

A clean single-platform reference for NVIDIA Jetson: what each module is, how much AI performance it delivers, which software stack it runs, and how to turn a trained model into a fast on-device TensorRT engine. The single most important thing to get right is the Super Mode TOPS numbers — since JetPack 6.2 (Feb 2025) every Orin Nano / Orin NX TOPS figure changed, and quoting the pre-Super numbers is the most common stale answer.

Sources: NVIDIA official Jetson pages and docs — the Jetson Orin product page, the JetPack 6.2 Super Mode blog, JetPack SDK pages, and the Jetson Product Lifecycle. Every spec below is grounded in those; provenance is noted inline. Specs and "currently shipping" status change — defer to the live NVIDIA selector page when in doubt.

The one thing to get right: Super Mode changed the TOPS

Since JetPack 6.2 (Jetson Linux 36.4.3, Feb 2025), NVIDIA shipped Super Mode — a software-only clock/power boost for Orin Nano and Orin NX modules. It roughly doubles advertised INT8 throughput with no hardware change. When someone asks "how many TOPS is an Orin Nano", answer with the Super number and say it requires JetPack 6.2 + the right nvpmodel profile:

• Orin Nano 8GB / Orin Nano Super: 67 TOPS (was 40 pre-Super)
• Orin Nano 4GB: 34 TOPS
• Orin NX 8GB: 117 TOPS (was 70 pre-Super)
• Orin NX 16GB: 157 TOPS (was 100 pre-Super)

All TOPS figures NVIDIA quotes are INT8 Sparse; dense INT8 is half. Enable Super Mode after flashing JetPack 6.2 with sudo nvpmodel -m 2 (Nano) or sudo nvpmodel -m 0 (NX), then reboot.

Module cheat-sheet (specs → software stack)

Confirm the exact module first (cat /proc/device-tree/model), then read the row. AGX Orin and Xavier NX do not have Super Mode — their numbers are fixed.

Module	GPU arch	AI perf (INT8 Sparse)	Memory	Max JetPack	Notes
AGX Orin 64GB	Ampere (2048-core)	275 TOPS	64GB LPDDR5	JetPack 6.x	Flagship; no Super Mode (already MAXN)
AGX Orin 32GB	Ampere (1792-core)	248 TOPS	32GB LPDDR5	JetPack 6.x
Orin NX 16GB	Ampere (1024-core)	157 TOPS (Super)	16GB LPDDR5	JetPack 6.x	Was 100 pre-Super
Orin NX 8GB	Ampere (1024-core)	117 TOPS (Super)	8GB LPDDR5	JetPack 6.x	Was 70 pre-Super
Orin Nano 8GB	Ampere (1024-core)	67 TOPS (Super)	8GB LPDDR5	JetPack 6.x	"Orin Nano Super" devkit; was 40
Orin Nano 4GB	Ampere (512-core)	34 TOPS (Super)	4GB LPDDR5	JetPack 6.x
Xavier NX	Volta (384-core)	21 TOPS	8 / 16GB LPDDR4x	JetPack 5.1.x (last)	No JetPack 6; Volta, not Ampere
Jetson Nano (legacy)	Maxwell (128-core)	472 GFLOPS FP16	2 / 4GB LPDDR4	JetPack 4.6.x (EOL)	NO INT8 tensor accel — do NOT quote a TOPS number

Two facts that are easy to get wrong:

• The original Jetson Nano has no INT8 tensor cores. Its number is 472 GFLOPS FP16 (128-core Maxwell), not a TOPS figure. JetPack 4 reached End of Life in Nov 2024 (last release 4.6.6 / L4T R32.7.6); JetPack 5/6 do not support it.
• Xavier NX tops out at JetPack 5.1.x (TensorRT 8.5.x). It does not run JetPack 6.

JetPack / Jetson Linux (L4T) stack

JetPack is the board SDK. It bundles Jetson Linux (a.k.a. L4T — Linux for Tegra) plus CUDA, cuDNN, TensorRT, VPI, and DLA. One JetPack version maps to one L4T version:

JetPack	Jetson Linux (L4T)	Ubuntu	CUDA	TensorRT	Supports
6.2 (current prod)	36.4.3 (kernel 5.15)	22.04	12.6	10.3	Orin (Nano/NX/AGX) — Super Mode
6.0 / 6.1	36.3 / 36.4	22.04	12.x	10.x	Orin
5.1.x	35.x	20.04	11.4	8.5.x	Orin + Xavier
4.6.x (EOL)	32.7.x	18.04	10.2	8.2	Nano + Xavier (legacy)

• Check installed version on-device: sudo apt-cache show nvidia-jetpack or read /etc/nv_tegra_release for the L4T line.
• Container images come from nvcr.io (NGC). The image tag's L4T base must match the host JetPack major/minor — a JetPack 6 container will not run on a JetPack 5 host.

Workflows

Workflow 1 — Identify the module and its software stack

1. Module model: cat /proc/device-tree/model (e.g. "NVIDIA Jetson Orin Nano Developer Kit").
2. JetPack / L4T: sudo apt-cache show nvidia-jetpack | head (JetPack) and cat /etc/nv_tegra_release (L4T R-number).
3. Read the cheat-sheet row for AI perf, max JetPack, and whether Super Mode applies.
4. If the user expects JetPack 6 on Xavier NX or legacy Nano, stop — those modules cap at 5.1.x / 4.6.x respectively.

Workflow 2 — First setup, Super Mode, and clocks

Use when: 刚拿到 Jetson, 怎么开始, 刷机, 性能上不去, Super 模式怎么开.

1. Flash — use NVIDIA SDK Manager (host x86 Ubuntu) for full SDK, or the SD-card image for Orin Nano / Xavier NX devkits from the Jetson Download Center. The Orin Nano Super devkit needs JetPack 6.2 (or the *-super.conf config) to unlock Super Mode.
2. First boot — Ubuntu OEM setup (user, locale, network). The devkit boots to a full Ubuntu desktop.
3. Enable Super Mode (Orin Nano/NX on JetPack 6.2): sudo nvpmodel -m 2 (Nano) or sudo nvpmodel -m 0 (NX) → reboot. List profiles with sudo nvpmodel -q.
4. Pin max clocks — sudo jetson_clocks locks the GPU/CPU/EMC to max for benchmarking (it does NOT raise the nvpmodel cap; set the power mode first).
5. Verify — tegrastats shows live GPU%, EMC, power, and thermals; install jtop (sudo pip3 install jetson-stats) for a top-like dashboard.

Workflow 3 — Run a model with TensorRT (.onnx → .engine)

Use when: 怎么跑模型, TensorRT 转换, trtexec, engine, 推理慢.

On Jetson you do not ship a .pt/.onnx to production — you build a TensorRT engine for the target module. Engines are device + TensorRT-version specific: an engine built on one JetPack is not portable to another, so build on the target Jetson.

1. Export ONNX on your training host (torch.onnx.export, opset matching your TensorRT version).
2. Build the engine on the Jetson with trtexec (ships in /usr/src/tensorrt/bin/):

   /usr/src/tensorrt/bin/trtexec --onnx=model.onnx --saveEngine=model.engine --fp16
   

   Add --int8 (with calibration) for max throughput on Orin's INT8 path; --useDLACore=0 --allowGPUFallback to offload to the DLA accelerator and free the GPU.
3. Set the power mode first (Workflow 2) — benchmarking before nvpmodel/jetson_clocks gives misleadingly low FPS.
4. Profile — trtexec --loadEngine=model.engine reports latency/throughput; tegrastats confirms the GPU is actually loaded.

For prebuilt model pipelines, point users at NVIDIA jetson-inference (Hello AI World) and DeepStream rather than hand-rolling.

Worked examples

Example 1 — "Orin Nano 到底多少 TOPS?" Answer with the Super number and the caveat: "Orin Nano 8GB(含 Super 开发套件)是 67 TOPS(INT8 Sparse),4GB 版是 34 TOPS。这是 JetPack 6.2 Super Mode 下的数字 —— 早期资料写的 40 TOPS 是旧值。开启方式:刷 JetPack 6.2 后 sudo nvpmodel -m 2 再重启。Dense INT8 减半(约 33 TOPS)。"

Example 2 — "我的 Jetson Nano 能升到 JetPack 6 吗?" "不能。原版 Jetson Nano(Tegra X1 / Maxwell)最高只到 JetPack 4.6.x,且 JetPack 4 已于 2024 年 11 月 EOL。JetPack 5/6 只支持 Orin 和 Xavier。Nano 也没有 INT8 张量加速,算力是 472 GFLOPS FP16,不要按 TOPS 报。" If they need a newer stack, that's a hardware upgrade to Orin → route board-selection questions to rdk-ecosystem only if they're also comparing against RDK.

Example 3 — "怎么在 Orin NX 上把我的 onnx 跑起来,而且要快?" Walk Workflow 3: "先开性能模式 sudo nvpmodel -m 0(NX Super)+ sudo jetson_clocks,再在这块 NX 上用 trtexec 构建 engine:/usr/src/tensorrt/bin/trtexec --onnx=model.onnx --saveEngine=model.engine --fp16,要更快就上 --int8(需校准)或 --useDLACore=0 卸到 DLA。engine 与设备/TensorRT 版本绑定,必须在目标板上构建。用 tegrastats 确认 GPU 真的在跑。"

Example 4 — "Jetson 和 RDK X5 哪个适合我?" This is a cross-platform selection question — route to rdk-ecosystem, which owns RDK-vs-Jetson comparison anchored on the RDK side. Provide Jetson specs from the cheat-sheet here as input, but don't make the buying call in this skill.

Common pitfalls

❌ Don't	✅ Do
Quote Orin Nano as 40 TOPS / Orin NX as 100 TOPS	Use Super numbers (67 / 117 / 157) and name JetPack 6.2
Give the Jetson Nano a TOPS figure	Say 472 GFLOPS FP16 — no INT8 tensor accel
Tell a Xavier NX / legacy Nano user to install JetPack 6	Cap at JetPack 5.1.x (Xavier) / 4.6.x (Nano, EOL)
Copy a .engine between boards / JetPack versions	Build the TensorRT engine on the target Jetson
Benchmark before setting power mode	nvpmodel + jetson_clocks first, then measure
Run a JetPack-6 nvcr.io container on a JetPack-5 host	Match the container's L4T base to the host JetPack
Make the RDK-vs-Jetson buying call here	Route cross-platform selection to rdk-ecosystem

Reference map

Read this	When
module-specs.md	Full per-module spec sheet, Super Mode power profiles, TOPS (sparse/dense), lifecycle/EOL details
jetpack-stack.md	JetPack ↔ L4T ↔ CUDA/TensorRT version matrix, flashing methods, nvcr.io container matching, version-check commands
tensorrt-workflow.md	trtexec flags, ONNX export, INT8/DLA, jetson-inference / DeepStream entry points, profiling
scripts/jetson_lookup.py	Deterministic module → AI perf / GPU arch / max JetPack / Super-mode lookup