SPIKING Neural Networks (SNNs) are becoming an increasingly widespread tool for enhancing the energy efficiency of near-sensor execution of Artificial Intelligence (AI) tasks. These networks operate on event-based data, significantly reducing computational workload and power consumption. Neuromorphic processors efficiently exploit event sparsity; however, they are not yet considered mainstream in the market due to costs and availability.
An attractive alternative technology for executing SNNs is provided by FPGA devices. These offer flexible memory, DSP capabilities, and reconfigurable logic, which are useful for mapping use-case-specific blocks for spike encoding and decoding, i.e., converting the input-output of the task into spikes and vice versa. Within the EdgeAI project, the team from University of Cagliari (Sardinia), in short UniCA, has developed SYNtzulu, a Spiking Neural Network (SNN) processing system designed to harness the computational efficiency of SNN algorithms on low-cost, low-power, and highly accessible FPGAs.
SYNtzulu (Pronounced: [sìn-tzu-lu], meaning “mosquito” or “gnat” in the Sardinian language) is designed to be lightweight and achieves several advantages:
- It enables time-series data analysis on FPGA-based SNN accelerators, bridging the gap between edge sensor processing and state-of-the-art SNN processors on FPGAs.
- It combines the simplicity of clock-driven execution with the potential to exploit sparsity. Part of the reconfigurable logic is dedicated to implementing a step-by-step activity tracking mechanism, reducing synapse dynamics processing time when spikes are sparse.
It integrates low-end FPGAs and lightweight SNN models to drastically reduce power consumption and comply with the power and energy requirements of deeply embedded processing systems.
Figura 1: Processing scheme in Syntzulu
Figura 2: Syntzulu architecture overview
Summary of Synzulu’s Key Features:
- Optimized for low-cost, low-power FPGA platforms.
- RISC-V subsystem for flexible runtime control.
- Computational capability: 192 million synapses and 48 million neurons per second.
- Ultra-low power consumption: peaks at 14.2 mW during inference and drops to just 0.3 mW in idle mode.
Data transfer: input via a 12 MHz SPI, output via a 2 MHz UART.
Application Cases: is lightweight accurate enough?
SYNtzulu has been rigorously evaluated, demonstrating state-of-the-art accuracy with minimal power consumption across iEEG, sEMG, and ECG applications. This makes it an excellent solution for real-time, low-power sensor data processing.
These three applications serve as representative examples of real-time, power-constrained scenarios, each varying in inference rate and data bandwidth requirements.
To train the SNNs, we employed the PyTorch package SLAYER (Spike LAYer Error Reassignment) within the LAVA software framework, executing the training process on a T4 GPU available through Google Colab.
- iEEG Use Case: This regression task utilized the Mean Squared Error (MSE) loss function, measuring the difference between neuron potentials in the network’s last layer and the target variables. Each output neuron potential corresponds to one of five target variables, with its firing condition disabled to prevent resetting the potential to zero.
- ECG and sEMG Use Cases: These classification tasks followed a different approach, where each output neuron was assigned to a specific class. The ECG model contained five output neurons, while the sEMG model featured thirteen. Training was conducted using SLAYER’s SpikeRate loss function, which conditions output neurons to fire at a specific rate when recognizing their respective classes.
Across all cases, the obtained accuracy meets or exceeds state-of-the-art benchmarks. This underscores our approach as a highly effective method for lightweight SNN applications, even in comparison to mainstream neuromorphic processors.
The video shows the encoded spikes deriving from a forearm-acquired EMG signal, and the corresponding decoded force of the five fingers, following accurately the dataset ground truth.
SYNtzulu is Open-Hardware!
The complete IP is available for download here: https://github.com/gianlucaleone/SYNtzulu.
To facilitate your experience, we’ve provided two Python scripts that streamline the processes of training, quantization, deployment, and testing of SNNs on the SYNtzulu platform. Dive into the repository, follow our step-by-step tutorial, and customize SYNtzulu to fit your specific needs!
We welcome your feedback, offer support if required, and are eager to explore potential collaborations!
Blog signed by: UNICA team