Zero-Copy GPU Inference from WebAssembly on Apple Silicon

A developer working on a project called Driftwood has demonstrated a breakthrough technique for zero-copy GPU inference directly from WebAssembly modules on Apple Silicon. The innovation leverages Apple's Unified Memory Architecture where CPU and GPU share the same physical memory, eliminating the traditional PCIe bus barrier. By creating a three-link chain—using mmap for page-aligned memory, Metal's zero-copy buffer API, and Wasmtime's custom memory allocator—data can flow between WebAssembly sandboxes and GPU compute without defensive copies or serialization overhead.

In traditional systems, moving data from WebAssembly's isolated linear memory to GPU acceleration requires two expensive copies: first out of the sandbox into host memory, then across the PCIe bus into GPU memory. Apple Silicon's architecture changes this physics entirely, allowing both CPU and GPU to read and write the same physical bytes. The developer validated the approach with a 128×128 matrix multiplication test where the WebAssembly module fills matrices, the GPU computes via GEMM shader, and results appear back in the module's memory with zero errors across 16,384 elements.

The technique enables a new runtime architecture where WebAssembly serves as the control plane and GPU as the compute plane with minimal overhead. This could revolutionize how AI inference systems are built, particularly for stateful applications where latency and memory efficiency are critical. While still early-stage research, the approach demonstrates what's possible when hardware architecture aligns with software abstraction layers to eliminate traditional bottlenecks.