Skip to content

Benchmark

Per-step wall time of the Lennard-Jones Langevin droplet example across droplet sizes, for the full cross-product of device (CPU / GPU), neighbour list and kernels (Warp / array (NumPy/CuPy) / JAX / C++). Lower is better. The neighbour-list backends are:

  • matscipy-neighbours — this library (matscipy_neighbours), CPU + GPU;
  • matscipy 1.2.0 — the classic matscipy package's neighbour_list, the CPU reference this library descends from;
  • vesin — vesin, CPU + GPU.

Test machine

  • CPU: Intel(R) Core(TM) Ultra 7 356H (16 logical cores)
  • GPU: NVIDIA RTX PRO 500 Blackwell Generation Laptop GPU

CPU threading

On the CPU the matscipy-neighbours list is benchmarked both single-threaded (OMP_NUM_THREADS=1, the (1t) rows) and multi-threaded (all 16 logical cores, the (mt) rows). The C++ force loop is OpenMP-parallel and follows the same setting; the Warp and array kernels and the JAX backend use their own threading. The classic matscipy 1.2.0 and vesin CPU lists are single-threaded. The GPU rows are unaffected.

Empty cells

vesin and matscipy 1.2.0 only feed the Warp and array kernels: JAX uses the dense neighbour_matrix and the C++ example uses the in-tree C++ core, so those rows are left empty. matscipy 1.2.0 is CPU-only, so it has no GPU rows. A blank in an otherwise-populated row marks a size that exceeded the GPU memory (e.g. the JAX dense matrix and the CuPy/vesin GPU paths at the largest sizes on this card).

Run configuration: reduced LJ units, cutoff 2.5, dt 0.005, friction 1.0, temperature 0.7; logarithmically spaced sizes (100 → 1000000 atoms). Up to 40 steps per point (fewer for the largest systems; JAX and Warp are compiled once during an untimed warm-up).

Time vs. number of atoms

Configuration 100 atoms 1000 atoms 10000 atoms 100000 atoms 1000000 atoms
Warp · matscipy-neighbours · GPU 0.85 1.42 5.32 31.22 244.88
Warp · matscipy-neighbours · CPU (mt) 0.31 3.77 13.88 91.11 1093.46
Warp · matscipy-neighbours · CPU (1t) 0.15 1.97 25.04 256.83 2807.42
Warp · matscipy 1.2.0 · CPU 0.19 3.19 41.62 461.17 5314.42
Warp · vesin · GPU 1.66 3.60 11.53 275.64 5426.82
Warp · vesin · CPU 0.37 3.78 45.86 568.37 7309.84
array (NumPy/CuPy) · matscipy-neighbours · GPU 2.75 7.27 49.40 477.16 5858.19
array (NumPy/CuPy) · matscipy-neighbours · CPU (mt) 0.87 6.11 39.11 518.21 5851.86
array (NumPy/CuPy) · matscipy-neighbours · CPU (1t) 0.32 4.02 54.21 723.14 7660.85
array (NumPy/CuPy) · matscipy 1.2.0 · CPU 0.47 5.49 74.91 907.25 10138.81
array (NumPy/CuPy) · vesin · GPU 4.28 9.49 72.26 946.87 —
array (NumPy/CuPy) · vesin · CPU 0.49 6.96 84.99 1054.04 12846.20
JAX · matscipy-neighbours · GPU 5.10 6.48 10.74 69.57 —
JAX · matscipy-neighbours · CPU (mt) 3.30 6.63 19.96 324.35 3042.44
JAX · matscipy-neighbours · CPU (1t) 2.16 4.91 30.27 501.67 4883.77
JAX · matscipy 1.2.0 · CPU — — — — —
JAX · vesin · GPU — — — — —
JAX · vesin · CPU — — — — —
C++ · matscipy-neighbours · GPU 1.03 1.54 3.83 27.39 287.22
C++ · matscipy-neighbours · CPU (mt) 0.83 6.95 21.49 73.30 612.11
C++ · matscipy-neighbours · CPU (1t) 0.08 1.88 23.79 240.28 2384.93
C++ · matscipy 1.2.0 · CPU — — — — —
C++ · vesin · GPU — — — — —
C++ · vesin · CPU — — — — —

(values are ms/step)

How to read it:

  • The neighbour-list build dominates the step, so the list choice drives the scaling: matscipy-neighbours' cell list stays close to linear on both devices, the classic matscipy 1.2.0 list is a single-threaded CPU reference, and vesin's GPU path falls behind for these large, low-density droplets.
  • The kernel choice mostly shifts the curve: the fused C++/CUDA and Warp kernels avoid materialising per-pair arrays, the array (NumPy/CuPy) path is the simplest, and JAX jit-compiles a dense masked sum.
  • On the CPU, the matscipy-neighbours (mt) rows pull away from (1t) as the system grows; and even single-threaded, matscipy-neighbours (1t) is already faster than the classic matscipy 1.2.0 and vesin CPU lists.

This page is generated by examples/lj_langevin/benchmark.py. Regenerate it on your own hardware with:

python examples/lj_langevin/benchmark.py --build build --doc-out docs/benchmark.md

For the C++ rows, build with -DBUILD_EXAMPLES=ON (and -DENABLE_CUDA=ON for the GPU binary); the other rows need pip install jax warp-lang vesin muTimer matscipy==1.2.0 in the interpreter that runs this driver.