[Paper Review] Low-Power Low-Latency Keyword Spotting and Adaptive Control with a SpiNNaker 2 Prototype and Comparison with Loihi

Yexin Yan, Terrence C. Stewart+8 more|arXiv (Cornell University)|Sep 18, 2020

Advanced Memory and Neural Computing44 references4 citations

TL;DR

This paper evaluates the SpiNNaker 2 neuromorphic prototype against Loihi in two benchmark tasks: keyword spotting and adaptive robotic control. It demonstrates that SpiNNaker 2's MAC array enables superior efficiency in high-dimensional vector-matrix multiplication, outperforming Loihi in DNN-based keyword spotting and high-dimensional adaptive control, while Loihi remains more efficient in low-dimensional cases.

ABSTRACT

We implemented two neural network based benchmark tasks on a prototype chip of the second-generation SpiNNaker (SpiNNaker 2) neuromorphic system: keyword spotting and adaptive robotic control. Keyword spotting is commonly used in smart speakers to listen for wake words, and adaptive control is used in robotic applications to adapt to unknown dynamics in an online fashion. We highlight the benefit of a multiply accumulate (MAC) array in the SpiNNaker 2 prototype which is ordinarily used in rate-based machine learning networks when employed in a neuromorphic, spiking context. In addition, the same benchmark tasks have been implemented on the Loihi neuromorphic chip, giving a side-by-side comparison regarding power consumption and computation time. While Loihi shows better efficiency when less complicated vector-matrix multiplication is involved, with the MAC array, the SpiNNaker 2 prototype shows better efficiency when high dimensional vector-matrix multiplication is involved.

Motivation & Objective

• To evaluate the performance of the SpiNNaker 2 neuromorphic prototype in real-world AI workloads.
• To compare SpiNNaker 2's MAC array-based computation with Loihi's dedicated spiking circuits in terms of power and latency.
• To assess the efficiency of hybrid DNN/SNN workloads on neuromorphic hardware with mixed-precision accelerators.
• To determine under which conditions SpiNNaker 2 outperforms Loihi in neural network inference and online learning tasks.
• To provide insights into the architectural trade-offs between general-purpose processing with accelerators (SpiNNaker 2) versus specialized spiking circuits (Loihi).

Proposed method

• Implemented a DNN-based keyword spotting network on the SpiNNaker 2 prototype using its MAC array for efficient vector-matrix multiplication.
• Ported the same keyword spotting task to Loihi using a spiking neural network (SNN) version, enabling direct comparison of DNN vs SNN inference.
• Implemented an adaptive control algorithm using SNNs on both platforms, with online synaptic weight updates based on error signals.
• Employed dynamic voltage and frequency scaling (DVFS) and adaptive body biasing (ABB) in SpiNNaker 2 to optimize energy efficiency.
• Used the same neural network architecture and hyperparameters across both platforms to ensure fair performance comparison.
• Measured computation time and active energy consumption for each benchmark task under varying input dimensions.

Experimental results

Research questions

• RQ1 How does the SpiNNaker 2 prototype’s MAC array affect inference efficiency in DNN-based keyword spotting compared to Loihi’s SNN implementation?
• RQ2 In what input dimension regimes does SpiNNaker 2 outperform Loihi in adaptive control tasks involving online learning?
• RQ3 How do the energy and latency characteristics of SpiNNaker 2 and Loihi compare when handling high-dimensional vector-matrix multiplication?
• RQ4 What are the performance trade-offs between general-purpose processing with accelerators (SpiNNaker 2) and dedicated spiking circuits (Loihi) in hybrid DNN/SNN workloads?
• RQ5 Can the MAC array in SpiNNaker 2 effectively accelerate both DNN and SNN workloads, and under which conditions does it provide a performance advantage?

Key findings

• The SpiNNaker 2 prototype achieved faster inference and higher energy efficiency than Loihi in the DNN-based keyword spotting task due to the use of its MAC array for vector-matrix multiplication.
• For adaptive control, SpiNNaker 2 outperformed Loihi in computation time and active energy consumption when the input dimension was high, where MAC array utilization was maximized.
• Loihi showed better efficiency than SpiNNaker 2 in low-dimensional adaptive control tasks, where the general-purpose Arm core in SpiNNaker 2 dominated the computation and consumed more energy.
• The MAC array in SpiNNaker 2 provided a significant speedup in vector-matrix multiplication compared to the software-emulated equivalent on SpiNNaker 1, with estimated improvements in clock cycles.
• The combination of a general-purpose Arm core and specialized MAC accelerators in SpiNNaker 2 enabled efficient execution of both DNN and SNN workloads, unlike Loihi which is optimized only for SNNs.
• The results indicate that MAC array-based acceleration is particularly beneficial for high-dimensional neural network workloads, making SpiNNaker 2 a strong candidate for such applications in embedded AI.