Endurance-Aware Mapping of Spiking Neural Networks to Neuromorphic Hardware

Abstract

Neuromorphic computing systems are embracing memristors to implement high density and low power synaptic storage as crossbar arrays in hardware. These systems are energy efficient in executing Spiking Neural Networks (SNNs). We observe that long bitlines and wordlines in a memristive crossbar are a major source of parasitic voltage drops, which create current asymmetry. Through circuit simulations, we show the significant endurance variation that results from this asymmetry. Therefore, if the critical memristors (ones with lower endurance) are overutilized, they may lead to a reduction of the crossbar's lifetime. We propose eSpine, a novel technique to improve lifetime by incorporating the endurance variation within each crossbar in mapping machine learning workloads, ensuring that synapses with higher activation are always implemented on memristors with higher endurance, and vice versa. eSpine works in two steps. First, it uses the Kernighan-Lin Graph Partitioning algorithm to partition a workload into clusters of neurons and synapses, where each cluster can fit in a crossbar. Second, it uses an instance of Particle Swarm Optimization (PSO) to map clusters to tiles, where the placement of synapses of a cluster to memristors of a crossbar is performed by analyzing their activation within the workload. We evaluate eSpine for a state-of-the-art neuromorphic hardware model with phase-change memory (PCM)-based memristors. Using 10 SNN workloads, we demonstrate a significant improvement in the effective lifetime.

On the Mitigation of Read Disturbances in Neuromorphic Inference Hardware

“…In a recent work, we have shown that the unit parasitic resistance of bitlines/wordlines increases from 1Ω at 65 nm to 3.8Ω at 16 nm [10]. Such increase in the value of parasitic resistance leads to a higher voltage applied across each OxRRAM cell in the crossbar, which further reduces its state transition time.…”

“…This is due to an increase in the voltage within the crossbar at scaled nodes. Second, the variation of state transition time increases at smaller nodes due to higher voltage and current variations [10]. Finally, the state transition time of OxRRAM cells also depends on the resistance state.…”

DFSynthesizer: Dataflow-based Synthesis of Spiking Neural Networks to Neuromorphic Hardware

“…This is obtained from the propagation delay of current through the synaptic elements in the crossbar. As shown in many recent works [99,100,102], the current propagation delay within a crossbar depends on the speciic synaptic elements that are being activated in the crossbar. This is due to the diference in the amount of parasitic components on the bitlines and wordlines of a crossbar along the diferent current paths.…”

“…DecomposeSNN [16] decomposes an SNN to improve the cluster utilization. There are also performance-oriented SNN mapping approaches such as [7,11,15,86], energy-aware SNN mapping approaches such as [101], circuit aging-aware SNN mapping approaches such as [10,67,84,88,91], endurance-aware SNN mapping approaches such as [93,99,102], and thermal-aware SNN mapping approaches such as [100]. These approaches are evaluated with emerging SNN based applications [9,31,43,50,64,75], which we also use to evaluate DFSynthesizer.…”

Energy-Efficient Respiratory Anomaly Detection in Premature Newborn Infants

“…Each cluster can then fit onto a tile of the hardware. Then, the clusters are mapped to the tiles to optimize one or more hardware metrics such as energy [47,48], latency [49][50][51][52][53], circuit aging [54][55][56][57][58][59], and endurance [60][61][62]. We use the energy-aware mapping technique of [48].…”