Zero-Copy GPU Inference from WebAssembly on Apple Silicon

Article

• Technique shares WebAssembly linear memory directly with GPU on Apple Silicon.
• Apple’s Unified Memory Architecture eliminates CPU-GPU copy overhead.
• Enables zero-copy, zero-serialization inference pipelines from WASM modules.
• Currently works only in wasmtime, not browsers.

Discussion

• fulafel: unified memory isn’t Apple-exclusive; x86 iGPUs and old Intel Macs worked similarly.
• jedisct1 and saagarjha questioned the point of WASM if it only runs in one headless runtime on one arch.
• nl: this is just “memory control in WASM works”—Apple Silicon details are noise.
• pjmlp: flagged security implications of bypassing WASM memory isolation.

Discuss on HN