NVIDIA Jetson 上兼顧散熱與功耗的推理

Thermal- and Power-Aware Inference on NVIDIA Jetson

Focus: Jetson deployment and performance tuning\
Audience: Flutter, Android, edge AI, and robotics engineers

The problem

A model that performs well during a short benchmark may throttle after
sustained operation. Power mode, ambient temperature, memory pressure,
and simultaneous ROS workloads all affect tail latency.

What this guide covers

This tutorial focuses on practical architecture and performance
engineering for production systems. The central principle is to measure
the complete pipeline, keep latency-sensitive work close to the
hardware, and make failure states visible to the operator.

Reference architecture

Workload scheduler → model runtime → telemetry collector → policy
controller. The Flutter UI reports operational status and diagnostics,
while performance policy runs on the Jetson and respects system-level
constraints.

Implementation plan

Establish a baseline using the exact Jetson module, carrier board, power supply, cooling, JetPack, and model build.
Measure warm-up and long-duration runs, not only a short inference loop.
Log temperature, power mode, clocks, utilization, memory, and p50/p95/p99 latency with timestamps.
Use documented power modes and cooling appropriate to the hardware; do not bypass thermal protections.
Apply adaptive sampling or model selection when workload exceeds a latency or thermal budget.

Example

# Start with system-supported monitoring tools for your installed image.\n# Check available options with: tegrastats --help\ntegrastats --interval 1000

The snippets are illustrative starting points, not drop-in integrations.
Match APIs and version requirements to the exact SDK, OS image, hardware
revision, and runtime in your project.

Performance checklist

Measure end-to-end latency, not only one function or model invocation.
Report p50 and p95/p99 latency alongside throughput, memory use, and dropped work.
Bound queues and define what happens when consumers fall behind.
Avoid repeated frame copies and expensive work on Flutter's UI isolate.
Test sustained workloads after the device reaches thermal equilibrium.
Keep UI state, device state, and robot state distinct.
Add timeouts, cancellation, reconnection, and explicit stale-data indicators.
Profile on the actual target hardware using a reproducible build.

Common pitfalls

Do not compare benchmarks collected under different power modes or cooling conditions.
Do not optimize average latency while ignoring p99 stalls.
Avoid changing clocks or power settings without understanding board limits and vendor guidance.