MiniMCTAD: Minimalist Monte Carlo Transport Architecture Design
Siqing Fu, Tiejun Li, Jianmin Zhang, Sheng Ma, Sheng Liu · 2021
Monte Carlo (MC) method is a kind of statistical experiment method, which has been widely used in scientific research, especially in solving particle transport problems. The increase in application scale requires faster machine speeds, so we propose a novel and simple architecture, MiniMCTAD, to speed up MC Transport applications. We improve parallelism by interating more cores on a single chip, rather than improving single-threaded performance through complex on-chip designs. To achieve high-performance and low-power architecture, we design the CPU pipeline structure, execution unit, memory hierarchy and explore the design space. The most commonly used architecture simulators, gem5 and McPAT, are employed to evaluate the appropriate architecture model. The experimental results show that the performance of MiniMCTAD architecture is 1.416x that of out-of-order core architecture, and it also has less area and lower power consumption. In that same number of cores, the MiniMCTAD architecture achieves a 4.45x advantage on performance per watt and 2.78x advantage on performance per area.