Skip to content
QuickMDSim

LAMMPS GPU (KOKKOS) on a cloud runner

Run LAMMPS on an NVIDIA L4 GPU in QuickMDSim (KOKKOS, Pro plan). Free and Basic stay on CPU. One MPI rank, billed at 12× a 1-core CPU second.

GPU (KOKKOS) is live on Pro. Free and Basic stay on CPU (LAMMPS 22 Jul 2025, stable update 5). Cloudflare Containers have no GPU — the GPU path is a separate CUDA + KOKKOS image on one L4.

What you pick

  • App: GPU · L4 next to CPU cores. Locked unless you are on Pro (or staff).
  • 1 MPI rank, 1 GPU. We inject -k on g 1 -sf kk. Do not add /kk suffixes.
  • Credits: same wallet. L4 burns 12 CPU-seconds per wall-second. 60 minute cap.
  • API / MCP: accelerator=gpu, gpu_sku=l4 only.

When GPU wins

Many-body styles with neighbor rebuilds on larger cells — Tersoff, EAM, LJ on tens of thousands of atoms. The SiC cascade starter is the honest demo. Tiny tutorials often lose to 4 CPU cores after GPU overhead.

What is not on GPU

  • Free and Basic plans
  • Classic LAMMPS GPU package
  • ReaxFF, MEAM, ML potentials (not compiled on either image)
  • Multi-GPU, multi-rank KOKKOS, T4
  • User CLI flags — no -k on from the input

Known-bad pair styles are rejected before a GPU starts. Unknown styles may still run, and may be slower than CPU.

CPU path

Free: 1 core. Paid: 2, 4, 8, or 16. Pro also: 32 or 64. Eight cores and above run on a separate machine, not on the GPU. Credits = wall seconds × requested cores.

Cite LAMMPS: Citing LAMMPS.