A Parallel SystemC Virtual Platform for Neuromorphic Architectures

Abstract

Simulating complex neuromorphic architectures and benchmarking demanding Artificial Neural Network (ANN) applications requires significant computational resources during the early design phase. This paper presents a novel parallel SystemC-based Virtual Platform (VP) specifically designed for RISC-V multicore systems integrating multiple computing-in-memory (CiM) neuromorphic accelerators. By exploring various VP segmentation architectures, the system exploits host multicore capabilities to achieve significant speedups compared to conventional sequential SystemC execution.

Report

Key Highlights

• Challenge Addressed: The difficulty of efficiently simulating and benchmarking complex neuromorphic architectures running ANN applications on traditional virtual platforms.
• Innovation: Introduction of a parallel SystemC-based Virtual Platform (VP).
• Target System: The VP is specifically designed to model RISC-V multicore platforms.
• Acceleration Focus: The simulated platform includes multiple computing-in-memory (CiM) neuromorphic accelerators.
• Performance Goal: Achieve faster simulation execution by effectively exploiting the multicore capabilities of the host system, demonstrating speedup over sequential SystemC.

Technical Details

• Core Technology: SystemC, utilized in a parallel execution model rather than the conventional sequential approach.
• Architectural Target: RISC-V multicore systems, indicating support for heterogeneous computing environments.
• Accelerator Type: Neuromorphic accelerators based on the Computing-in-Memory (CiM) paradigm, crucial for energy-efficient AI processing.
• Optimization Method: Exploration and comparison of different Virtual Platform (VP) segmentation architectures to maximize parallelism and distribute computational load across host cores.
• Application Focus: Enabling the simulation of artificial neural network applications for benchmarking and architecture comparison.

Implications

• Enabling Neuromorphic Development: Provides the necessary high-performance simulation tools to accelerate the design and validation of specialized neuromorphic hardware, a key area of emerging computing.
• RISC-V Heterogeneity: Directly supports the growing RISC-V ecosystem by offering a platform to simulate complex heterogeneous designs that couple RISC-V cores with specialized AI accelerators early in the development cycle.
• Reduced Time-to-Market: Faster simulation times mean designers can conduct more exhaustive architecture design exploration, leading to optimized hardware designs and reduced development cycles for embedded neuromorphic systems.
• Benchmarking Standard: Establishes a methodology for computationally feasible benchmarking of various competing architectures under real-world ANN application workloads.