Domain-Specific Acceleration of Gravity Forward Modeling via Hardware-Software Co-Design.

Area of Science:

• Geophysics
• Scientific Computing
• Hardware Acceleration

Background:

• Gravity forward modeling is crucial for geophysics but computationally intensive.
• Traditional CPU-based methods achieve gains through algorithmic optimization.
• Field-programmable gate arrays (FPGAs) offer acceleration potential but face programmability and nonlinear function challenges.

Purpose of the Study:

• To develop an FPGA-based co-processor for accelerating gravity forward modeling.
• To address the challenges of programmability and nonlinear function implementation on FPGAs.

Main Methods:

• Integrated a RISC-V core with a custom instruction set for key computations.
• Implemented dynamic task scheduling across eight pipelined processing units for high parallelism.
• Utilized a piecewise linear approximation method optimized via stochastic gradient descent (SGD) for nonlinear operations.

Main Results:

• Achieved up to 179x speedup compared to a high-end CPU at 250 MHz.
• Improved energy efficiency by 2040x.
• Demonstrated high parallelism and retained programmability on the AMD UltraScale+ ZCU102 FPGA.

Conclusions:

• The proposed FPGA co-processor significantly accelerates gravity forward modeling.
• This work presents the first known FPGA-based acceleration design for gravity forward modeling.
• The approach offers a viable solution for compute-intensive geophysical simulations.