AutomataNexus LLCWhitepaper · May 2026
HVAC Prediction

Enki —
Unrolled 8-Step GRU Controller

Enki is the FCOG facility mechroom GRU controller. Named for the Sumerian god of water, knowledge, and crafts. A lighter-weight alternative to Enlil using GRU gates (reset + update) instead of LSTM gates (input + forget + cell + output), achieving 50,583 FPS — over 120x faster than the LSTM variant — at the cost of slightly reduced temporal memory depth.

Author
Andrew Jewell Sr.
Organization
AutomataNexus LLC
Framework
AxonML (Rust)
Silicon
Hailo-8
Front matterAbstract

Abstract

Background

Enki is the FCOG facility mechroom GRU controller. Named for the Sumerian god of water, knowledge, and crafts. A lighter-weight alternative to Enlil using GRU gates (reset + update) instead of LSTM gates (input + forget + cell + output), achieving 50,583 FPS — over 120x faster than the LSTM variant — at the cost of slightly reduced temporal memory depth.

Approach

The model was trained in AxonML (a pure-Rust deep learning framework) and compiled through the Hailo Dataflow Compiler (DFC 3.33.1) targeting Hailo-8 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-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289)), Enki achieves 50,582 FPS (hw_only) with 0.057 ms hardware latency at 0.96 W average power draw.

Conclusion

Enki 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 hvac prediction application.

ModelEnki
DomainHVAC Prediction
ArchitectureUnrolled 8-Step GRU Controller
Target siliconHailo-8
Measured onHailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289)
DFC compiler3.33.1
FrameworkAxonML v0.6 (pure-Rust, CUDA + CPU backends)
AuthorAndrew Jewell Sr. · ORCID 0009-0005-2158-7060
OrganizationAutomataNexus LLC · Fort Wayne, Indiana
Part I · Model01

Executive overview

Enki is the FCOG facility mechroom GRU controller. Named for the Sumerian god of water, knowledge, and crafts. A lighter-weight alternative to Enlil using GRU gates (reset + update) instead of LSTM gates (input + forget + cell + output), achieving 50,583 FPS — over 120x faster than the LSTM variant — at the cost of slightly reduced temporal memory depth.

50,582 FPS
Throughput
0.057 ms
HW Latency
0.96 W
Power (avg)
8
Target

Network I/O

Input: 8-step sensor sequence [1, 1, 8, 57]. Output: next-state prediction scalar.

Part I · Model02

Architecture

Unrolled 8-Step GRU Controller

8-timestep unrolled GRU with input_size=57, hidden_size=57. Each timestep: reset gate (sigmoid), update gate (sigmoid), candidate hidden state (tanh), and gated update. 6 weight matrices per timestep (3 input projections + 3 recurrent projections). The GRU recurrence pattern `h = (1-z)*n + z*h_prev` is fully unrolled into element-wise multiply and add operations that map directly to NPU element-wise units.

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).

Part II · Silicon03

Silicon performance

Measured on production hardware via hailortcli benchmark with 5-second sustained inference. Device: Hailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289).

MetricMeasured Value
FPS (hw_only)50,582.50
FPS (streaming)35,657.40
HW Latency0.057000 ms
Power (streaming avg)0.96500 W
Power (streaming max)0.98300 W
Power (idle)0.74979 W
QuantizationINT8 (post-training, DFC calibration)
DFC Compiler3.33.1
HailoRT4.20.0
Measured OnHailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289)
 Table 03-1 — Production silicon measurements, 5s sustained inference.
Part II · Silicon04

Deployment

Deployed as a single HEF binary. No ONNX runtime, TensorFlow Lite, or Python inference stack required at the edge.

Target siliconHailo-8
Measured onHailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289)
DFC compiler3.33.1
QuantizationINT8 (post-training, production telemetry calibration)
RuntimeHailoRT (vendor runtime)
Edge platformRaspberry 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-8 dataflow engine over PCIe. Inference begins immediately with deterministic per-frame latency. No model conversion, graph optimization, or warmup phase required.

Back matter05

References

  1. Jewell, A. (2026). AxonML: A Pure-Rust Deep Learning Framework for Edge Inference. AutomataNexus LLC. Technical whitepaper.
  2. Hailo Technologies Ltd. (2024). Hailo Dataflow Compiler User Guide. DFC v3.33.1.
  3. Hailo Technologies Ltd. (2024). Hailo-8 Product Datasheet.

AutomataNexus LLC · Fort Wayne, Indiana · andrew.jewellsr@automatanexus.com
Andrew Jewell Sr. · ORCID 0009-0005-2158-7060
May 2026 · All rights reserved.