What's Next

You can now write the full GPU surface: residency bindings, forall and gpu fn kernels, on-device reduction, shared-memory tiling, atomics, warp shuffles, vectors, and interactive gpu frame programs that ship to the browser.

Coming next on the GPU roadmap:

• Async GPU — async gpu blocks and explicit streams, overlapping transfer with compute.
• Kernel debugging — a gpu simulate mode: CPU interpretation of kernels with race and divergence detectors, plus in-kernel printing.
• Zero-copy on unified memory — opt-in host views over device buffers on UMA hardware.
• Native backends — SPIR-V, PTX, and Metal for the features WGSL core can't express: 64-bit atomics, cooperative matrix, bf16/fp8.
• Richer reductions — multi-workgroup hierarchical reduce, min/max and general closure folds.