Odin v0 Performance Benchmarks
Odin v0 benchmarks a Jetson Orin Nano Super paired with the Axelera Metis M.2 D-IMC accelerator against a standalone NVIDIA Jetson Orin Nano Super across three satellite-relevant workloads.
High-Level Results
| Benchmark | Metric | Odin v0 | NVIDIA Jetson | Delta |
|---|---|---|---|---|
| Pose Estimation | Inference latency | 0.82 ms | 6.4 ms | 7.8× faster |
| Pose Estimation | Timing jitter (σ) | 0.40 ms | 17.44 ms | 43.6× tighter |
| Pose Estimation | System power | 10.3 W | 14.8 W | 30.4% lower |
| Cloud Detection | Inference power | 10.1 W | 14.4 W | 29.9% lower |
| Cloud Detection | Scene processing | 29.88 s | 60.34 s | 2.1× faster |
| GEMM (N=1024) | Efficiency | 3,390 GOPS/W | 852 GOPS/W | 3.98× more efficient |
Test Environment
| Component | Value |
|---|---|
| Platform | NVIDIA Jetson Orin Nano Super 8GB |
| Accelerator | Axelera Metis M.2 (PCIe 3.0 x4) |
| OS | Ubuntu 22.04 (JetPack 6.2.1) |
| Baseline runtime | TensorRT 10.3 / CUDA 12.6 |
| Accelerator SDK | Voyager SDK v1.5 |
| Precision | Odin v0: INT8 · Jetson: FP16 or INT8 (per test) |
| Power monitoring | INA260 external current sensor |
Methodology
Models are exported to ONNX opset 17, then compiled separately for each backend: trtexec for TensorRT and the Voyager Compiler for the D-IMC accelerator. Clocks are pinned via nvpmodel and jetson_clocks before each run. Each configuration includes a warmup phase followed by a timed measurement loop.
Efficiency (GOPS/W) is measured as net active power — device inference minus device idle baseline — ensuring a fair comparison of compute efficiency independent of platform idle draw.
Workloads
1. GEMM Raw Throughput
NxN INT8 matrix multiplication (N = 64–1024). Establishes raw compute efficiency and the operating point where D-IMC spatial execution outperforms GPU temporal execution.
2. Spacecraft Pose Estimation
Pose-ResNet50 (4.1 GFLOPs, 224×224 input) for 6-DoF spacecraft pose from monocular video frames. Characterizes inference latency, jitter, and system power for a real-time GNC pipeline.
3. Sentinel-2 Cloud Detection
DTACSNet-CD (U-Net/MobileNetV2, 0.62 GFLOPs) over full 10,980 × 10,980 px Sentinel-2 scenes (2,809 tiles at 224×224). Characterizes tiled-inference throughput and power under sustained load.