Date: 2026-11-11 · Plan phase: 0
Create the project following the dev guide, confirm PyTorch trains on the local GPU, and measure the two workloads the plan depends on so that the compute split between CPU and GPU is decided from data.
uv init --lib --python 3.12 sso-scimlunderC:\Users\tntth\Code\dev, branch renamed tomain.pyproject.toml: PyTorch pulled from the CUDA 13.0 wheel index on Windows and from the CPU index elsewhere, so CI does not download CUDA libraries..gitignoreextended for datasets, checkpoints, and experiment tracking.scripts/check_gpu.py: device report plus two timings.tests/test_environment.py: import check, a known-spectrum eigenvalue check in float64, CPU against GPU agreement..github/workflows/ci.yml: lint and CPU tests on every push.docs/PLAN.md.
cd C:\Users\tntth\Code\dev
uv init --lib --python 3.12 sso-sciml
cd sso-sciml
git branch -M main
uv sync
uv run scripts/check_gpu.py
uv run ruff check
uv run pytestOutput of uv run scripts/check_gpu.py:
python 3.12.13 (Windows 11)
torch 2.14.1+cu130
cuda available=True build=13.0
device NVIDIA GeForce RTX 3070 capability=8.6
memory 6.95 GiB free of 8.00 GiB
threads 8 CPU threads
batched eigvals, float64 (seconds per call)
batch n cpu cuda cpu/cuda
4096 20 0.148 8.550 0.02
4096 40 0.610 16.506 0.04
1024 100 1.311 29.719 0.04
MLP training throughput (samples per second)
cpu float32 116,337
cuda float32 2,406,917
cuda AMP 3,104,608
peak GPU memory allocated: 160 MiB
- Training: the GPU with mixed precision is about 27 times faster than the CPU on a 4×512 MLP at batch 8192. Mixed precision adds about 29 % over float32 on the GPU.
- Eigenvalues: the CPU solves 4,096 matrices of size 20 in 0.15 s. The GPU takes 8.6 s for the same call.
The expectation going in was that batched eigenvalue solves would be a GPU workload. They are not: the GPU is 23 to 58 times slower than the CPU across the three sizes tested. The non-symmetric eigenvalue routine does not run as a batched GPU kernel for small matrices, so the call pays per-matrix overhead. The plan was changed before it was written down: eigenvalue-based data generation is assigned to CPU processes.
Not yet tested: whether batched fixed-step time integration is faster on the GPU. That benchmark is in Phase 3.
- PyTorch is the single array framework. JAX has no GPU support on native Windows.
- Solver code in float64; surrogate training in float32 with mixed precision.
- Eigenvalue solves on the CPU. GPU for training, inference, and (pending the Phase 3 benchmark) batched time integration.
- Python pinned to 3.12.
- MIT licence.
Phase 1, step 1: write docs/EQUATIONS.md for the dq-frame network and the First Benchmark Model.