Flexi-NeurA: A Configurable Neuromorphic Accelerator with Adaptive Bit-Precision Exploration for Edge SNNs
Researchers' configurable neuromorphic accelerator achieves 1.1ms inference latency while using only 1,623 logic cells.
A research team from the University of Tehran has unveiled Flexi-NeurA, a breakthrough configurable neuromorphic accelerator designed specifically for running Spiking Neural Networks (SNNs) on power-constrained edge devices. The core innovation addresses a critical bottleneck in edge AI: most existing neuromorphic hardware platforms are rigid, offering limited adaptability to diverse workloads and design goals like balancing accuracy, latency, and power consumption.
The Flexi-NeurA architecture provides unprecedented design-time configurability, allowing engineers to customize neuron models, network structures, and—crucially—bit-precision settings for key parameters like synaptic weights and membrane potentials. This flexibility is powered by a companion tool called Flex-plorer, a heuristic-guided Design Space Exploration (DSE) engine. Flex-plorer automatically finds the most cost-effective fixed-point precision configurations based on user-defined trade-offs between accuracy and hardware resource usage (like logic cells and memory), then generates the corresponding RTL code.
In comprehensive evaluations, the framework demonstrated remarkable efficiency. A three-layer fully connected network with Leaky Integrate-and-Fire (LIF) neurons, mapped onto just two Flexi-NeurA processing cores, achieved 97.23% accuracy on the MNIST dataset. This was accomplished with an ultra-low inference latency of 1.1 milliseconds, while consuming a mere 111 milliwatts of total power and utilizing only 1,623 logic cells and 7 Block RAMs on an FPGA. This performance establishes Flexi-NeurA not just as a research prototype, but as a highly scalable and practical platform for bringing advanced, efficient neuromorphic computing to real-world edge intelligence applications, from sensors to drones.
- Achieves 97.23% MNIST accuracy with ultra-low 1.1ms latency and 111mW total power consumption.
- Uses a heuristic-guided tool (Flex-plorer) to auto-optimize fixed-point precision for weights and potentials, balancing accuracy and hardware cost.
- Highly configurable architecture allows customization of neuron models and network structures, generating tailored RTL code for specific edge applications.
Why It Matters
Enables complex, efficient AI directly on sensors and drones, reducing cloud dependency and latency for real-time edge intelligence.