/jetson-diagnostic

Reading tegrastats output

tegrastats is the canonical Jetson telemetry tool. Output tokens and spelling change across Jetson Linux / L4T releases and SoCs (for example older guides document EMC X%@Y while newer Orin/Thor guides document EMC_FREQ with several alternate shapes). Treat any single cheat-sheet as approximate.

There is no single NVIDIA table that covers every minor Jetson Linux revision. The Reported Statistics glossary can differ between nearby releases (for example R38.2.x vs R38.4.0). Prefer the archive whose Jetson Linux version string best matches cat /etc/nv_tegra_release on the device; if NVIDIA has not published that exact slug, use the closest published archive for your product line and still sanity-check against live tegrastats output.

Official documentation (NVIDIA)

The authoritative field glossary and CLI options live in the Jetson Linux Developer Guide, topic Tegrastats Utility (under Applications and Tools → Jetson Linux Development Tools). The HTML basename is stable:

DeveloperGuide/AT/JetsonLinuxDevelopmentTools/TegrastatsUtility.html

URL pattern

Substitute the archive folder name (Jetson Linux release) for <SLUG>:

https://docs.nvidia.com/jetson/archives/<SLUG>/DeveloperGuide/AT/JetsonLinuxDevelopmentTools/TegrastatsUtility.html

Browse NVIDIA’s Jetson documentation archives to find the <SLUG> that matches your image (for example r38.4, r36.4.4). Not every minor gets its own folder (for example r38.2.2 and r38.4.0 may 404 while r38.2.1 and r38.4 exist); use the nearest published slug.

Per-archive Tegrastats Utility pages (known-good slugs)

The links below were spot-checked (HTTP 200); NVIDIA may add or retire archive folders over time — if one fails, use the archives index and the URL pattern above.

Welcome indices (overview only, not the field table): r38.4, r36.5, r35.6.4.

Welcome-only context: Jetson Linux 38.2 documentation explicitly states that release does not support the Jetson Orin product family at GA time; use a 36.x guide for Orin-focused tegrastats wording when that applies to your stack (r38.2 Welcome).

Man page?

Many images ship tegrastats without a man page; use tegrastats --help on the device for a short option list, and the Tegrastats Utility topic for full semantics. NVIDIA also documents the binary location as /core/utils/tegrastats in the BSP package layout.

Very old Jetson Linux (e.g. L4T 32.x)

Archived L4T help bundles (such as Jetson Linux 32.7.6) may use an older web help shell; if search fails, use the PDF/HTML Development Guide for that release from the Jetson Download Center or cross-check a nearby r35.x archive link above.

Example line (typical Jetson Linux 36.x / 38.x style, Orin-class)

Not guaranteed character-for-character on every release; compare to your device and the version-matched Tegrastats Utility table.

RAM 4011/8138MB (lfb 8x4MB) SWAP 0/4068MB (cached 0MB) CPU [12%@1190,7%@1190,...] EMC_FREQ 0% GR3D_FREQ 35% AO@47C CPU@52C GPU@49C tboard@45C tdiode@46C VDD_GPU_SOC 1234mW VDD_CPU_CV 567mW

Recorded samples (optional; cite Jetson Linux revision)

Real captures help humans and agents pattern-match timestamps, *_FREQ @MHz spellings, Thor GR3D_FREQ three-tuples, and thermal zone names that the generic example above does not show. Treat samples as illustrative: a future Jetson Linux update can change formatting even on the same SKU.

Jetson AGX Thor Developer Kit — Jetson Linux R38 (release), REVISION: 4.0

Identifying lines from /etc/nv_tegra_release:

# R38 (release), REVISION: 4.0, GCID: 43443517, BOARD: generic, EABI: aarch64, DATE: Wed Dec 31 00:15:19 UTC 2025
# KERNEL_VARIANT: oot

Official field glossary for this line: Tegrastats Utility — Jetson Linux 38.4 archive.

tegrastats foreground sample (~idle system; includes leading timestamp as printed by this build):

04-21-2026 15:50:52 RAM 105265/125772MB (lfb 34x4MB) CPU [0%@972,0%@972,0%@972,0%@972,0%@972,0%@972,1%@972,0%@972,0%@972,0%@972,0%@972,8%@972,0%@972,0%@972] EMC_FREQ 0%@2750 GR3D_FREQ @[314,314,314] NVENC0_FREQ @314 NVENC1_FREQ @314 NVDEC0_FREQ @314 NVDEC1_FREQ @314 NVJPG0_FREQ @314 VIC off OFA_FREQ @315 PVA0_FREQ off APE 300 cpu@37.343C tj@38.875C soc012@37.468C gpu@38.875C soc345@37.406C VDD_GPU 1962mW/1962mW VDD_CPU_SOC_MSS 5887mW/5887mW VIN_SYS_5V0 5635mW/5635mW VIN 19780mW/19780mW

Notes for this capture: GR3D_FREQ shows three frequencies (three GPCs on Thor). No NVDLA* tokens — consistent with NVIDIA’s statement that NVDLA tegrastats reporting applies to AGX Orin, not Thor, on the same guide page.

Field-by-field (common tokens; names drift by release)

Quick triage

• GR3D_FREQ near 100% sustained → workload is GPU-bound.
• EMC / EMC_FREQ near 100% with low GPU → memory-bandwidth-bound; quantization or KV-cache reuse helps (token name depends on Jetson Linux version).
• CPU cores pinned at 100% but GPU low → likely Python overhead, decoder bottleneck, or pre/post-processing on CPU.
• AO/GPU above 90 °C sustained → thermal throttling imminent; check the cooling solution and nvpmodel.
• lfb showing only small blocks → memory fragmentation; large CUDA allocations may fail before total RAM is exhausted.

Useful invocations

# One sample, one second (stdout).
tegrastats --interval 1000 | head -n 1

# 30-second log to disk.
tegrastats --interval 1000 --logfile /tmp/tegra.log &
sleep 30
tegrastats --stop

# Stop any background tegrastats (preferred):
tegrastats --stop

If the device policy denies access to some counters, re-run with elevated privileges only as needed (same flags).

SKU notes

• Orin Nano / NX / AGX: layout above; no nvidia-smi.
• Thor: layout above is preserved; nvidia-smi is also available and gives a complementary GPU-centric view.