Fast Loosely-Timed Deep Neural Network Models with Accurate Memory Contention

The emergence of data-intensive applications, such as Deep Neural Networks (DNN), exacerbates the well-known memory bottleneck in computer systems and demands early attention design flow. Electronic System-Level (ESL) using SystemC Transaction Level Modeling (TLM) enables effective performance estimation, space exploration, gradual refinement. However, contention is often only detectable after detailed TLM-2.0 approximately-timed or cycle-accurate RTL models are developed. A detected at a late stage can severely limit available choices even require costly redesign. In this work, we propose novel loosely-timed contention-aware (LT-CA) modeling style that offers high-speed simulation close to traditional (LT) models, yet shows same accuracy for low-level (AT) models. Thus, our proposed LT-CA breaks speed/accuracy tradeoff between regular LT AT fast accurate observation visualization contention. Our extensible model generator automatically produces desired TLM-1 from DNN architecture description exploration focusing on We demonstrate approach with real-world industry-strength application, GoogLeNet. experimental results show 46x faster than equivalent an average error less 1% simulated time. Early detection contentions also suggests local memories computing cores eliminate applications.