Dendrite

GPU-accelerated CUDA kernels for lithium-ion battery simulation.

What these kernels do:

• Spherical diffusion - Lithium transport inside electrode particles (fast charging bottleneck)
• Butler-Volmer - Electrochemical kinetics at particle surfaces
• 2D thermal - Heat transport for thermal management

133x faster than NumPy. 197x faster than naive GPU code.

Why

The battery simulation community is stuck on CPU. Existing tools:

Tool	Backend	Notes
PyBaMM	CPU (CasADi/IDAKLU)	Excellent physics, no GPU
BattMo	CPU (MATLAB)	No GPU
COMSOL	CPU + cuDSS	2-5x GPU speedup (generic sparse solver)

We wrote hand-tuned CUDA kernels. Result:

Method	Time (10K particles, 100s sim)	Speedup
NumPy vectorized	351 ms	1x
CuPy naive	518 ms	0.7x (slower!)
Dendrite	2.6 ms	133x

Naive GPU ports don't help. Hand-tuned kernels do.

Build

make            # builds lib/libdendrite.{a,so}
make examples   # builds bin/simple_diffusion, bin/battery_particle
make benchmarks # builds bin/benchmark

Requires CUDA Toolkit 11.0+ and an NVIDIA GPU.

Usage

#include "dendrite.h"

// 2D diffusion (89% of peak bandwidth on RTX 3090)
dendrite_diffusion_2d(c_in, c_out, D, dx, dy, dt, nx, ny, stream);

// Butler-Volmer kinetics
dendrite_butler_volmer(eta, i0, j, alpha, T, n, stream);

// Spherical particle diffusion (for SPM-style models)
dendrite_spherical_diffusion(c, j_surf, D_s, R_p, dt, nr, n_particles, stream);

Performance

RTX 3090 (936 GB/s theoretical peak):

Kernel	Bandwidth	Peak %
2D Diffusion (4K x 4K)	832 GB/s	89%
Butler-Volmer (4M pts)	794 GB/s	85%
Spherical Diffusion	712 GB/s	76%

Docs

See docs/ for:

• DEVELOPMENT.md - Build setup, benchmarking methodology, LLM instructions
• BENCHMARKS.md - Full benchmark results
• OPTIMIZATION.md - Performance tuning findings
• PHYSICS.md - Battery physics background

License

MIT