Skip to content

GPU CUDA Acceleration Engine (v3.1 Future Prospect)

This document details the technical architecture for offloading continuous spatial field solvers—such as 2D/3D reaction-diffusion PDEs and airborne VOC advection—from CPU Numba JIT loops to dedicated GPU hardware via PyTorch tensor operators and custom CUDA C++ kernels.

1. Architectural Motivation & Performance Bottleneck

While Numba @njit JIT loops achieve sub-millisecond execution times for 2D grids up to \(512 \times 512\) (\(\sim 0.85\text{ ms/tick}\)), scaling to massive biotope grids (\(2048 \times 2048\) to \(4096 \times 4096\)) or 3D vertical canopy volumes (\(W \times H \times Z\), \(Z=16\)) introduces significant CPU memory bandwidth pressure.

flowchart TD
    CPU_Loop["Numba CPU Loop<br><i>2D/3D Stencil Iteration</i>"] -->|Memory Bandwidth Cap| CPU_Stall["CPU Cache Misses<br><i>Grid Sizes > 1024x1024</i>"]

    GPU_Offload["GPU CUDA Engine<br><i>PyTorch Tensor Parallelism</i>"] -->|Zero Copy Zero Latency| GPU_Speed["Massive Grid Execution<br><i>< 0.20 ms/tick @ 4096x4096</i>"]

2. Technical Architecture & Hybrid Execution Model

To avoid host-to-device memory transfer overhead (\(O(N)\) PCIe bus latencies), PHIDS employs a Hybrid GPU/CPU Execution Model:

  1. GPU Field Solvers: 2D/3D signaling matrices (GridEnvironment layers) reside permanently in GPU VRAM as 32-bit floating-point PyTorch Tensors.
  2. CUDA Stencil Kernels: Gaussian diffusion and semi-Lagrangian wind advection passes execute directly on GPU CUDA cores via shared-memory \(3 \times 3\) (or \(3 \times 3 \times 3\)) stencil blocks.
  3. Zero-Copy Interoperability: ECS entity interaction passes (metabolic burn, bites) access co-located GPU tensor buffers via Cupy or PyTorch CUDA pointers without CPU roundtrips.

3. Implementation Roadmap & Milestones

  • Phase 3.1.1: PyTorch tensorized 2D PDE solver replacing Numba array loops (phids/engine/systems/pde_gpu.py).
  • Phase 3.1.2: Custom C++/CUDA extension for 3D vertical canopy advection with height-dependent shear.
  • Phase 3.1.3: Zero-copy telemetry streaming directly from GPU VRAM into Zarr storage buffers via PyTorch-Zarr GPU connectors.