FPGA-based distributed computing microarchitecture for complex physical dynamics investigation

Area of Science:

• Computational physics
• Applied mathematics
• Hardware acceleration

Background:

• Partial differential equations (PDEs) are fundamental in fields like solid state physics, nuclear physics, and plasma physics.
• Solving complex PDEs often requires significant computational resources and time.
• Existing distributed systems face challenges in scalability and efficiency for certain PDE types.

Purpose of the Study:

• To introduce DCMARK, a novel distributed computing system designed for efficient PDE solving.
• To leverage the cellular neural network (CNN) paradigm for parallel integration of differential equations.
• To implement and test DCMARK on a Field-Programmable Gate Array (FPGA) for high reconfigurability and performance.

Main Methods:

• Developed DCMARK architecture based on the CNN paradigm, mapping one processor per equation.
• Implemented DCMARK on a single FPGA, optimizing processor design for minimal hardware requirements and interconnectivity.
• Created a 200-cell Korteweg-de Vries (KdV) equation solver to test the platform, comparing results with a high-performance PC.

Main Results:

• DCMARK demonstrated constant computation time, independent of the number of dynamical elements (cells) in the CNN array.
• The FPGA implementation achieved significant reductions in elaboration time compared to similar systems.
• A compact, reconfigurable system managed by a softcore processor facilitated efficient data/control communication with a PC Host.

Conclusions:

• DCMARK offers a highly efficient and reconfigurable solution for solving PDEs, particularly for 1-, 2-, and 3-D locally interconnected dynamical systems.
• The system's constant computation time provides a performance advantage over existing methods for large-scale simulations.
• The FPGA-based approach enables reduced elaboration times and enhanced flexibility for scientific investigations.