SCPN GPU Roadmap

The GPU acceleration program is tracked in docs/GPU_ACCELERATION_ROADMAP.md. This page summarizes the execution sequence and acceptance criteria for quick reference in the Sphinx docs set.

Scope

  • Target path: hybrid CPU/GPU execution for Grad-Shafranov + transport kernels

  • Explicitly non-speculative: implementation-first roadmap only

  • Tracking issue: GitHub #17

Execution Phases

  1. wgpu SOR kernel (red-black stencil + deterministic CPU fallback)

  2. GPU-backed GMRES preconditioning (CUDA/ROCm adapters with CPU fallback)

  3. Full multigrid on device (smooth/restrict/prolong + coupled nonlinear path)

Performance Targets

  • Phase 1: - 65x65: 2x to 4x - 257x257: 5x to 12x

  • Phase 2: - 2x to 6x speedup on linear-algebra-heavy inverse solves

  • Phase 3: - under 1 ms for control-loop grids - 10x to 30x speedup for 257x257+ workloads

Acceptance Gates

  • Correctness: residual behavior matches CPU reference within configured tolerance

  • Performance: measured speedups meet declared minimum floors

  • Operations: runtime capability detection + automatic CPU fallback

Reference

  • Full plan: docs/GPU_ACCELERATION_ROADMAP.md