The same 3D Euler solver on WebGPU - 128^3 Taylor-Green in seconds, validated against the CPU engine.
See it run - a worked example, 100% in this browser tab
The problem
Interactive 3D hydrodynamics at 128^3 is out of reach on a CPU in the browser, where a full Taylor-Green run takes roughly fifteen minutes.
The local-first solution
This plugin runs the same unsplit HLLC + MUSCL + RK3-TVD scheme f32-resident on WebGPU in the browser, validated cell-for-cell against the f64 CPU engine, bringing a full 128^3 Taylor-Green run down to seconds.
What it does
f32-resident HLLC + MUSCL + RK3-TVD on WebGPU
Measured ~3.9 ms/step at 128^3 (a full TGV run ~9 s native vs ~15 min on CPU)
Validated cell-for-cell vs the f64 CPU engine (advect L1 matches to 5 digits)
Taylor-Green energy conserved with enstrophy growth on the GPU path
Pure WebGPU: errors honestly with no adapter, never silently runs the slow CPU path
Honest scope
The GPU evolution is f32 and the trust label carries _gpu_f32 so the precision class is explicit and never implies f64. It is inviscid (ILES), a single-GPU prototype bridge to HPC, not production high-Reynolds DNS and not a replacement for Athena++/PLUTO/FLASH. The engine's own GeoNum verdict on the f64-upcast read-back is passed through.
Authorities cited
Taylor, G. I. & Green, A. E. (1937). Mechanism of the production of small eddies from large ones. Proc. R. Soc. Lond. A 158, 499-521.
Brachet, M. E. et al. (1983). Small-scale structure of the Taylor-Green vortex. J. Fluid Mech. 130, 411-452.
Toro, E. F. (1999). Riemann Solvers and Numerical Methods for Fluid Dynamics, 2nd ed. Springer. (HLLC, chapter 10.)
Run 128^3 on the GPU
Run the GPU solver in the browser and save the result to Sandbox, attach it to a Worklog case, or route it into a Gate client portal. Nothing leaves your machine to anyone's cloud.