Spikenaut-SNN-v2

A 16-neuron Leaky-Integrate-and-Fire (LIF) Spiking Neural Network trained on live cryptocurrency telemetry mining, high-frequency trading, and blockchain sync node telemetry. Designed for Xilinx Artix-7 FPGA deployment at 97 mW.

Architecture

Spec	Value
Neuron model	Leaky-Integrate-and-Fire (LIF)
Neurons	16
Input channels	16
Weight format	Q8.8 fixed-point
Learning rules	E-prop, OTTT, reward-modulated STDP
Clock	1 kHz (1ms resolution)
Training speed	35 us/tick
Memory footprint	1.6 KB
FPGA power	97 mW (25 mW dynamic, 72 mW static)
FPGA target	Xilinx Artix-7 xc7a35tcpg236-1 (Basys3)

16-Channel Input Map

Channels	Data Source	Function
0-1	DNX (Dynex)	PoUW solver health and neural baselines
2-3	Quai	Live on-chain reflex and sync confidence
4-5	Qubic	Epoch and tick cadence monitoring
6-7	Kaspa	High-frequency DAG settlement tracking
8-9	XMR (Monero)	Node stability and CPU L3 cache contention
10-11	Ocean	Data liquidity and staking prep
12-13	Verus	CPU-heavy validator tracking (AVX-512)
14-15	Thermal	Pain receptors -- power and temperature

Channels 14-15 are the network's pain receptors. When the GPU crosses 85C, the SNN receives negative reward and learns to avoid states that could damage the hardware.

This is bound to change in the future, not sure how I am going to change the input format.

Merged v2 Parameters

This model ships with a merged parameter set combining the best of three training sources:

Parameter	Source	Values
Thresholds (16)	Real trained weights	Graduated 1.125 to 1.594 per neuron
Decay rates (16)	Converted parameters	Graduated 0.80 to 0.95 per neuron
Hidden weights (256)	Real trained weights	Range 0.75 to 1.04, 76 unique values
Output weights (48)	Real trained weights	Signed: -0.164 to +0.258 (inhibitory + excitatory)

Q8.8 Fixed-Point Format

All .mem files use Q8.8 fixed-point encoding. Each line is one 4-digit hex value:

Hex: 0100  →  Decimal: 256  →  Float: 256/256 = 1.0
Hex: 00DA  →  Decimal: 218  →  Float: 218/256 = 0.852
Hex: 00CC  →  Decimal: 204  →  Float: 204/256 = 0.797

Negative values use two's complement: FFF9 = -0.027.

Files

Merged v2

dataset/merged_v2/
├── parameters.mem              # 16 neuron thresholds (Q8.8 hex)
├── parameters_decay.mem        # 16 decay rates (Q8.8 hex)
├── parameters_weights.mem      # 16x16 weight matrix (Q8.8 hex)
├── parameters_output_weights.mem # Output layer weights (signed Q8.8)
└── snn_model.json              # Full model definition (float values)

Each contains: parameters.mem, parameters_weights.mem, parameters_decay.mem

Verilog

// Load thresholds from Q8.8 hex file
reg [15:0] threshold_ram [0:15];
initial $readmemh("dataset/merged_v2/parameters.mem", threshold_ram);

// Load weights from Q8.8 hex file
reg [15:0] weight_ram [0:255];
initial $readmemh("dataset/merged_v2/parameters_weights.mem", weight_ram);

Training Results

Metric	Value
Architecture	Julia-Rust hybrid
Algorithm	E-prop + OTTT
Convergence	20 epochs
Training speed	35 us/tick
IPC overhead	0.8 us
Memory usage	1.6 KB
Training date	2026-03-22
Data sources	Kaspa mainnet, Monero mainnet

Known Limitations

• Monotonic hidden weight pattern: Root cause identified (2026-07-12) — NOT an export bug. Caused by degenerate training convergence: 8 uniform training samples + no inhibitory connections + identical E-prop/OTTT gradients. Fix: retrain with qubic_ticks_snn.jsonl (27K records) + add 4 inhibitory neurons (80:20 E:I ratio) + implement K-WTA sparsity.

• Purely excitatory hidden layer: All 256 hidden weights are positive. The network lacks inhibitory connections (negative weights) and recurrent feedback, which limits its capacity for noise suppression and temporal memory. A future training run should add ~4 inhibitory neurons and recurrent connections.

Hardware Baseline

Component	Spec
CPU	AMD Ryzen 9 9950X
GPU	NVIDIA RTX 5080 (Blackwell SM_120)
FPGA	Digilent Basys3 (Xilinx Artix-7 xc7a35tcpg236-1)
FPGA Power	97 mW total
FPGA LUTs	1,063 / 20,800 (5.11%)
FPGA Registers	1,091 / 41,600 (2.62%)
Timing WNS	3.727 ns (37.27% margin)
OS	Fedora 44

The Story

In 2013, a severe concussion left me unable to process the world's data the way I used to. Without access to neuro-rehabilitation, I decided to research on my own, and I started building what would become Spikenaut -- a neuromorphic system that learns from the raw signals of the machines I run every day. Originally inspired by the bottlenecks of my local GPU (RTX 5080), spent loads of money just to find out that I can't run massive LLM's on it for AI tutoring. So naturally my curious mind went on the internet to find alternatives. That is where I found Spiking Neural Networks, a low power alternative to traditional neural networks.

Unfortunately, the neuromorphic field is still in its early stages, and Spikenaut is just the beginning. I created Limen-Neural a GitHub organization over my experimental work in Neuromorphic computing. Meantime I have been modularizing all my work into reusable components in Limen-Neural. Feel free to check it out use the code to your liking, copy and use it in your own projects or use git dependencies. I'm still far from where I want it to be but I can guarantee you in a near future I will be there with benchmarks, docs with wiki and performance improvements.

As of the right now the weights are a mess, merged_v2 is where I am going to continue improving, the rest are more artifacts than anything. So expect updates over the time for new and improve SNN weights.

The name comes from "spike" (neural firing) and "naut" (navigator). This model is the brain -- the trained neural weights that turn raw telemetry into decisions.

Related

• Limen-Neural — github.com/Limen-Neural (runtime, FPGA export, training crates)
• Telemetry — rmems/Spikenaut-SNN-Telemetry
• Q8.8 export — silicon-bridge

License

Dual-licensed under MIT and Apache-2.0. Developed independently by Raul Montoya Cardenas