Platform · Benchmarks
Performance benchmarks
Independent benchmark data comparing Neurmorph SNN inference against NVIDIA Jetson Xavier NX and ARM Cortex-M7 baselines across representative edge sensor workloads. All tests run in controlled lab conditions; methodology and reproduction scripts available on request.
Top-line results
Methodology
How we benchmark
Controlled conditions
All devices tested at nominal clock speeds, ambient temperature 23°C, no other processes running. Power measured at the device rail using a calibrated current shunt at 10 µA resolution.
Standardized workloads
Five workloads: IMU gesture classification, 8-class audio keyword detection, vibration anomaly, thermal point detection, and mixed-sensor fusion. All models trained on public datasets.
Reproducible scripts
Benchmark suite open-sourced under Apache 2.0. Reproduction requires an NT3000 eval kit and a Jetson Xavier NX. Contact us for access to the benchmark harness repository.
Detailed results
Neurmorph NT3000 vs. Jetson Xavier NX
Workload: IMU 9-axis gesture classification, 12-class, 100-sample window. Model: 3-layer SNN (78K parameters). Accuracy measured on 2,000-sample holdout set.
| Metric | Neurmorph NT3000 | Jetson Xavier NX | Cortex-M7 @ 480 MHz |
|---|---|---|---|
| Inference / Watt | 47,200 | 693 | 1,840 |
| Latency median | 0.11 ms | 2.3 ms | 18.7 ms |
| Latency p99 | 0.34 ms | 4.1 ms | 24.2 ms |
| Active power draw | 38 mW | 9,800 mW | 240 mW |
| Idle power draw | 3.8 mW | 1,200 mW | 18 mW |
| Accuracy vs. dense baseline | -1.4% | 0.0% | -0.8% (TFLite int8) |
| Binary size | 87 KB | 12.4 MB (TensorRT) | 182 KB (TFLite) |
| SRAM requirement | 96 KB | 2 GB | 320 KB |
Neurmorph NT3000 eval kit firmware v1.2.0. Jetson Xavier NX tested with TensorRT 8.6 FP16 precision. Cortex-M7 tested with TFLite Micro int8. All latency values include I/O overhead at the sensor bus. Accuracy delta measured relative to float32 PyTorch dense baseline on identical holdout set.
Workload comparison
Power efficiency across workloads
Neurmorph inference/Watt advantage over Jetson Xavier NX (FP16 TensorRT) at matched accuracy ±2%. Higher is better.
| Workload | Efficiency gain | Latency (NT3000) | Accuracy delta |
|---|---|---|---|
| IMU gesture (12-class) | 68× | 0.11 ms | -1.4% |
| Audio keyword (8-class) | 54× | 0.28 ms | -2.0% |
| Vibration anomaly detection | 83× | 0.07 ms | -0.6% |
| Thermal point classification | 61× | 0.19 ms | -1.9% |
| Mixed-sensor fusion | 47× | 0.41 ms | -2.1% |
All workloads compiled with NMC v0.9.1, threshold calibration enabled, targeting NT3000 silicon. Contact [email protected] for the full benchmark dataset and reproduction instructions.
Test on your workload
Results on standardized workloads may not reflect your specific use case. Apply for a pilot to benchmark against your own dataset.