Seismic processing algorithms and inverse problems in general are computationally demanding and often executed on GPU-accelerated HPC systems. Moreover, recent approaches based on AI and machine learning show similar features and can significantly benefit from reduced-precision arithmetic.
The goal of the project is to investigate mixed-precision computing strategies for large-scale seismic processing and AI/ML workflows, exploiting modern GPU architectures and Tensor Cores. Different floating-point formats (fp32, fp16, bfloat16) will be combined to improve performance, energy efficiency and memory usage while preserving numerical stability and solution quality. The work will focus on identifying performance-critical kernels, analyzing their sensitivity to reduced precision, and designing hybrid-precision schemes where higher precision is retained only where necessary.