Abstract
Advances in neuroscience uncover the mechanisms employed by the brain to efficiently solve complex learning tasks with very limited resources. However, the efficiency is often lost when one tries to port these findings to a silicon substrate, since brain-inspired algorithms often make extensive use of complex functions such as random number generators, that are expensive to compute on standard general purpose hardware. The prototype chip of the 2nd generation SpiNNaker system is designed to overcome this problem. Low-power ARM processors equipped with a random number generator and an exponential function accelerator enable the efficient execution of brain-inspired algorithms. We implement the recently introduced reward-based synaptic sampling model that employs structural plasticity to learn a function or task. The numerical simulation of the model requires to update the synapse variables in each time step including an explorative random term. To the best of our knowledge, this is the most complex synapse model implemented so far on the SpiNNaker system. By making efficient use of the hardware accelerators and numerical optimizations the computation time of one plasticity update is reduced by a factor of 2. This, combined with fitting the model into to the local SRAM, leads to 62% energy reduction compared to the case without accelerators and the use of external DRAM. The model implementation is integrated into the SpiNNaker software framework allowing for scalability onto larger systems. The hardware-software system presented in this work paves the way for power-efficient mobile and biomedical applications with biologically plausible brain-inspired algorithms.
Introduction
The human brain is incredibly efficient at learning complex tasks with limited power, largely because its connections are sparse, noisy, and constantly rewiring—a process known as structural plasticity. Neuroscientists model this using "synaptic sampling," where synapses undergo a continuous random walk to explore new network architectures. However, simulating these brain-inspired algorithms on standard silicon often results in a loss of efficiency. These models require massive amounts of random number generation (RNG) and complex exponential functions that standard CPUs and GPUs struggle to process efficiently. This paper presents a hardware-software co-design approach using the SpiNNaker 2 prototype to make these noisy, dynamic brain models computationally viable.
Methods and Hardware Architecture
The researchers utilized a 4-core prototype of the SpiNNaker 2 system. Unlike its predecessor, SpiNNaker 2 features ARM Cortex-M4F cores augmented with specific hardware accelerators for RNG and exponential functions.
To map the synaptic sampling algorithm to this hardware, the authors employed several key techniques:
- Hardware Acceleration: Expensive software-based Gaussian noise generation was replaced with hardware-accelerated uniform noise, and exponential calculations were offloaded to a custom fixed-point accelerator.
- Local SRAM Computation: To avoid the energy and latency penalties of fetching data from off-chip DRAM, the model was compressed using 16-bit floating-point data types so all synapse parameters could reside entirely in the chip's local SRAM.
- Random Reallocation of Synapses: The authors introduced a novel software trick: when a synapse's weight drops to zero, its memory slot is immediately recycled and reassigned to a new random target neuron. This keeps the fanout constant and accelerates learning by exploring the parameter space more aggressively.
Key Results
The system was evaluated using a custom, synthetic closed-loop reinforcement learning task where the network learned to route random spike patterns based on a global reward signal.
- Speedup: Using hardware accelerators reduced the computation time for a single plasticity update by more than half (from 236 to 110 clock cycles). The time spent on RNG and exponentials plummeted from 62% to just 18% of the total workload.
- Energy Efficiency: By keeping all data in local SRAM and utilizing the accelerators, energy consumption per time step was slashed by 62% (from 450.3 μJ down to 171 μJ).
- Real-Time Capability: The system successfully ran the task in real-time, computing a 1ms biological time step in only 0.76ms.
Conclusion and Future Work
This work demonstrates that standard ARM cores paired with specific math accelerators provide a perfect blend of software flexibility and hardware efficiency for neuromorphic applications. While the current prototype is compute-bound at approximately 3,000 plastic synapses per core, the authors plan to scale this implementation to the full 144-core SpiNNaker 2 chip. Future applications include processing high-dimensional data from event-based cameras for mobile and biomedical use cases.
Summary updated