Microarchitecture Characterization and Analysis of Emerging Neuro-Symbolic AI Workloads

The emergence of neuro-symbolic A.I. (NSAI) workloads introduces a new and powerful approach to artificial intelligence by combining deep learning techniques with traditional rules-based approaches. Dedicated research to better understand and enhance these NSAI models is needed as they are emerging as a major field in AI. Examples of NSAI models analyzed in this paper are the Neural Logic Machine (NLM) and the Logic Tensor Network (LTN). Given that neuro-symbolic models are still an evolving research area, the performance characteristics remain insufficiently understood. Conducting an analysis of the performance characteristics of NSAI workloads offers valuable insights into potential avenues for accelerating and enhancing the performance of these models. By examining the microarchitecture performance characteristics of the NSAI models, including compute workload, memory workload, schedule statistics, warp state statistics, and executed instruction mix, a deeper understanding of the internal mechanisms of these NSAI models can be gained. The obtained performance characteristics have the potential to facilitate improvements in LI and L2 cache hit rates, issue latency, execution latency, IMC and scoreboard stalls, and other microarchitecture opportunities. In this paper, we explore the promising acceleration design opportunities of the NLM and LTN through the analysis of their microarchitecture performance and characterization.