Mnemosyne (Aegis) —
Multi-Scale Conv2D Face Encoder (Production)
Production deployment of the Mnemosyne face recognition encoder on Hailo-10H. The largest Aegis model, running at 3,126 FPS. At 47.5 degrees C average die temperature, it can run continuously without throttling on the Pi 5 platform.
Abstract
Background
Production deployment of the Mnemosyne face recognition encoder on Hailo-10H. The largest Aegis model, running at 3,126 FPS. At 47.5 degrees C average die temperature, it can run continuously without throttling on the Pi 5 platform.
Approach
The model was trained in AxonML (a pure-Rust deep learning framework) and compiled through the Hailo Dataflow Compiler (DFC 5.3.0) targeting Hailo-10H silicon. Post-training INT8 quantization was applied during the DFC compilation pass with production telemetry calibration data. The resulting Hailo Executable Format (HEF) binary executes on Hailo’s fixed-function dataflow architecture with deterministic latency and zero framework overhead at the edge.
Results
On production hardware (Hailo-10H (NexusWatch, FW 5.3.0)), Mnemosyne (Aegis) achieves 3,126 FPS with 47.5 °C average die temperature (max 47.9 °C).
Conclusion
Mnemosyne (Aegis) is production-ready as a single HEF binary deployed to edge devices with no external dependencies beyond the HailoRT vendor runtime. The model meets real-time latency requirements for its target biometric security application.
| Model | Mnemosyne (Aegis) |
| Domain | Biometric Security |
| Architecture | Multi-Scale Conv2D Face Encoder (Production) |
| Target silicon | Hailo-10H |
| Measured on | Hailo-10H (NexusWatch, FW 5.3.0) |
| DFC compiler | 5.3.0 |
| Framework | AxonML v0.6 (pure-Rust, CUDA + CPU backends) |
| Author | Andrew Jewell Sr. · ORCID 0009-0005-2158-7060 |
| Organization | AutomataNexus LLC · Fort Wayne, Indiana |
Executive overview
Production deployment of the Mnemosyne face recognition encoder on Hailo-10H. The largest Aegis model, running at 3,126 FPS. At 47.5 degrees C average die temperature, it can run continuously without throttling on the Pi 5 platform.
Network I/O
Input: RGB face crop [1, 3, 112, 112]. Output: 512-dim ArcFace embedding.
Architecture
Multi-Scale Conv2D Face Encoder (Production)
12-block multi-scale residual network with squeeze-and-excitation attention. 112x112 RGB input, 512-dim output. This is the most compute-intensive model in the biometric suite, benefiting most from Hailo-10H over H8 (3,126 FPS vs 4,588 on H8 — again H8 wins due to resource utilization, but H10 provides thermal headroom for sustained multi-model concurrent inference).
Compilation constraints
All AxonML models targeting Hailo silicon are compiled under the fixed-function dataflow constraints: no dynamic control flow, no variable-length dimensions, all activations representable in INT8 after calibration, and no operations requiring dedicated softmax hardware (replaced with ReLU gating or depthwise convolution equivalents where necessary).
Silicon performance
Measured on production hardware via hailortcli benchmark with 5-second sustained inference. Device: Hailo-10H (NexusWatch, FW 5.3.0).
| Metric | Measured Value |
|---|---|
| FPS (streaming) | 3,125.75 |
| Die Temperature (mean) | 47.52 °C |
| Die Temperature (min) | 46.07 °C |
| Die Temperature (max) | 47.89 °C |
| Quantization | INT8 (post-training, DFC calibration) |
| DFC Compiler | 5.3.0 |
| HailoRT | 5.3.0 |
| Measured On | Hailo-10H (NexusWatch, FW 5.3.0) |
Deployment
Deployed as a single HEF binary. No ONNX runtime, TensorFlow Lite, or Python inference stack required at the edge.
| Target silicon | Hailo-10H |
| Measured on | Hailo-10H (NexusWatch, FW 5.3.0) |
| DFC compiler | 5.3.0 |
| Quantization | INT8 (post-training, production telemetry calibration) |
| Runtime | HailoRT (vendor runtime) |
| Edge platform | Raspberry Pi 5 + Hailo AI HAT+ (M.2 Key M) |
Deployment procedure
Copy the .hef binary to the target device. hailortcli run loads the HEF directly into the Hailo-10H dataflow engine over PCIe. Inference begins immediately with deterministic per-frame latency. No model conversion, graph optimization, or warmup phase required.
References
- Jewell, A. (2026). AxonML: A Pure-Rust Deep Learning Framework for Edge Inference. AutomataNexus LLC. Technical whitepaper.
- Hailo Technologies Ltd. (2024). Hailo Dataflow Compiler User Guide. DFC v5.3.0.
- Hailo Technologies Ltd. (2024). Hailo-10H Product Datasheet.
AutomataNexus LLC · Fort Wayne, Indiana · andrew.jewellsr@automatanexus.com
Andrew Jewell Sr. · ORCID 0009-0005-2158-7060
May 2026 · All rights reserved.