Depth Anything 3 - Jetson Orin NX Benchmarks

February 5, 2026 ยท View on GitHub

Performance benchmarks for Depth Anything 3 (DA3) models running on NVIDIA Jetson Orin NX 16GB with TensorRT 10.3 optimization.

Test Date: February 2, 2026 Hardware: Jetson Orin NX 16GB (Seeed reComputer J4012), JetPack 6.2 TensorRT Version: 10.3 Precision: FP16


Executive Summary

ConfigurationFPSLatencyUse Case
DA3-Small @ 256x256110 FPS9.1msHigh-speed robotics
DA3-Small @ 308x30893 FPS10.9msReal-time robotics (recommended)
DA3-Small @ 400x40064 FPS15.8msBalanced quality/speed
DA3-Small @ 518x51840 FPS25.0msHigher quality
DA3-Base @ 518x51819 FPS51.4msQuality-focused
DA3-Large @ 518x5187.5 FPS132msOffline processing

Recommendation: For real-time robotics applications, use DA3-Small at 308x308 or 400x400 resolution for optimal speed/quality balance.


Resolution Benchmarks (DA3-Small)

Testing DA3-Small model at different input resolutions to find the optimal speed/quality tradeoff.

ResolutionThroughput (FPS)Latency (mean)Latency (p99)Engine Size
518x51840.0925.0ms25.6ms64MB
400x40063.5815.8ms16.4ms60MB
308x30892.5810.9ms11.1ms60MB
256x256110.209.1ms9.3ms56MB

Resolution Scaling Analysis

  • 2x speedup: Reducing from 518 to 400 yields 1.6x faster inference
  • 2.3x speedup: 308x308 provides 2.3x improvement over native resolution
  • 2.7x speedup: 256x256 achieves maximum throughput at 110 FPS

Quality Considerations

Lower resolutions reduce depth map detail but maintain relative depth accuracy. For navigation and obstacle avoidance, 308x308 provides sufficient spatial resolution while enabling real-time performance.


Model Size Benchmarks (518x518)

Comparing DA3 model variants at native 518x518 resolution.

ModelParametersThroughput (FPS)Latency (mean)Latency (p99)Engine Size
DA3-Small~24M40.0525.0ms25.6ms64MB
DA3-Base~97M19.2451.4ms52.8ms211MB
DA3-Large~335M7.51132.2ms134.2ms674MB

Model Scaling Analysis

  • Small to Base: 4x more parameters, 2x slower (19 vs 40 FPS)
  • Base to Large: 3.5x more parameters, 2.6x slower (7.5 vs 19 FPS)
  • Small to Large: 14x more parameters, 5.3x slower (7.5 vs 40 FPS)

Quality vs Speed Tradeoff

ModelRelative QualityRelative SpeedBest For
SmallGoodFastestReal-time robotics, embedded
BaseBetterModerateQuality-sensitive applications
LargeBestSlowestOffline processing, benchmarking

Memory Usage

GPU Memory (Estimated)

ModelResolutionGPU Memory
DA3-Small518x518~1.2GB
DA3-Small308x308~0.8GB
DA3-Base518x518~2.5GB
DA3-Large518x518~6GB

Disk Space (TRT Engines)

ConfigurationEngine Size
DA3-Small (all resolutions)~240MB total
DA3-Base (518x518)211MB
DA3-Large (518x518)674MB

Deployment Recommendations

Real-Time Robotics (30+ FPS required)

Recommended: DA3-Small @ 308x308 or 400x400

  • 308x308: 93 FPS with 11ms latency
  • 400x400: 64 FPS with 16ms latency
  • Both leave headroom for other processing

Quality-Focused Applications (15+ FPS acceptable)

Recommended: DA3-Small @ 518x518 or DA3-Base @ 518x518

  • Small: 40 FPS, good quality
  • Base: 19 FPS, better quality for detailed scenes

Offline/Batch Processing

Recommended: DA3-Large @ 518x518

  • 7.5 FPS suitable for non-real-time applications
  • Best depth quality for dataset generation

Test Methodology

Hardware Configuration

  • Device: NVIDIA Jetson Orin NX 16GB
  • JetPack: 6.2
  • Power Mode: MAXN (15W)
  • Cooling: Active fan cooling

Benchmark Parameters

  • Tool: trtexec (TensorRT benchmark utility)
  • Iterations: 100 inference passes
  • Warmup: 2000ms
  • Precision: FP16

Engine Build Settings

trtexec \
    --onnx=model.onnx \
    --saveEngine=model.engine \
    --fp16 \
    --memPoolSize=workspace:2048MiB \
    --optShapes=pixel_values:1x1x3xHxW

Reproducing Benchmarks

Prerequisites

  1. Jetson Orin NX with JetPack 6.2+
  2. TensorRT 10.3
  3. This repository deployed via deploy_jetson.sh

Running Resolution Benchmarks

cd ~/depth_anything_3_ros2
bash scripts/benchmark_resolutions.sh

Running Model Size Benchmarks

cd ~/depth_anything_3_ros2
bash scripts/benchmark_models.sh

Version History

DateChanges
2026-02-02Initial benchmarks on Jetson Orin NX 16GB
2026-02-02Added thermal/stability validation (10-min sustained load test)

Thermal/Stability Validation

10-Minute Sustained Load Test

Test performed with DA3-Small @ 518x518 under continuous inference load.

MetricValue
Duration600.06 seconds
StatusPASSED
Throughput40.79 FPS
Latency (mean)24.73ms
Latency (min)24.25ms
Latency (max)27.88ms
Latency (p99)25.19ms

Stability Analysis

  • FPS Stability: Maintained consistent 40.79 FPS throughout 10-minute test
  • Latency Variance: Only 3.63ms spread (min to max) indicating stable thermals
  • Thermal Throttling: None detected (performance remained constant)
  • p99 Latency: 25.19ms ensures predictable real-time behavior

Conclusions

The Jetson Orin NX 16GB demonstrates excellent thermal stability for sustained DA3 inference workloads. No performance degradation was observed over the 10-minute test period, making it suitable for continuous robotics applications.


References