subgroups

The subgroups feature exposes SIMD-wave level operations (ballot, shuffle, reductions like subgroupAdd) to WGSL compute and fragment shaders. Invocations in the same subgroup - typically 32 or 64 lanes, hardware dependent - can exchange values in registers, which is dramatically faster than staging data through workgroup memory with barriers.

It matters most for GPGPU workloads: reductions, scans, sorting, and machine-learning inference kernels often see large speedups from subgroup intrinsics. Engines also use it for light culling and occlusion pipelines. Because subgroup size varies by hardware (and can vary between pipelines on the same device), portable code must handle a range via subgroupMinSize/subgroupMaxSize in the adapter info.

Subgroups shipped in Chrome 125 for supported hardware after a long origin-trial period, and the companion subgroups-f16 capability folded into it. Older drivers, some mobile GPUs, and software adapters do not support it - which is exactly what the availability number on this page measures.