Had the opposite experience. Our JS FEM solver (~550ms per load case) was rewritten in Rust and dropped to ~270ms. But we compile to native.exe, not WASM — we call it via stdin/stdout with JSON from a Node.js compute engine. Tried the WASM route first but the serialization overhead for large stiffness matrices ate the gains, exactly like this article describes. Native binary + stdin/stdout turned out to be the sweet spot: no boundary tax, no FFI, and you get full native SIMD. The sparse solver variant (sprs crate, COO/CSC assembly) scales even better for larger models.