mirror of
https://github.com/JuliaFEM/JuliaFEM.jl.git
synced 2026-09-20 10:08:31 +00:00
Jukka Aho
caf60e4356
- Comprehensive guide for GPU-friendly FEM architecture benchmarks - State management strategy comparison: AoS vs SoA (coalesced memory access) - Matrix-free Newton-Krylov benchmarks with Anderson acceleration - Performance metrics: memory bandwidth (GB/s), execution time, iterations - Expected results: Strategy 2 (SoA) 10× faster on GPU, 500-900 GB/s bandwidth - Matrix-free + Anderson: 6-10× speedup, 2.5× fewer iterations - Hardware requirements and tested platforms (RTX 4090, 3090, 3080) - Troubleshooting guide: CUDA setup, OOM errors, slow CPU benchmarks - Performance expectation tables for 1M elements and 100K DOFs - Validation: correctness checks and convergence verification - 244 lines documenting GPU optimization strategies and benchmarking methodology