shader-f16

With shader-f16, WGSL code can declare f16 values, vectors, and matrices - after enable f16; - and run arithmetic natively at half precision. On hardware with real fp16 ALUs this roughly doubles throughput per register and halves memory traffic for buffers and varyings that opt in.

The big consumers are machine-learning inference (weights and activations in fp16), post-processing chains, and mobile renderers fighting bandwidth limits. Code must be written for it deliberately: f16 has ~3 decimal digits of precision and a max value of 65504, so naive ports can produce banding or NaNs.

Support requires both the adapter capability and enabling the feature at device creation. Desktop support is broad on modern GPUs; the gap you see in the availability data is mostly older hardware and software adapters.