v135 · javascript

Float16Array

Adds the Float16Array typed array. Number values are rounded to IEEE fp16 when writing into Float16Array instances.

concepts

  1. Float16Array round-trip

    Bit-level demo: type a value, see what survives an fp16 round-trip, and compare byte counts against Float32 / Float64.

  2. Image tensor downsample

    The motivating ML / graphics use case from the TC39 proposal. A 64×64 image is normalised into a tensor; the fp32 and fp16 reconstructions sit side-by-side along with bytes used and peak per-channel error.

  3. Half-precision budget

    Type any Float64 value and see what it becomes in Float32 vs Float16. Bytes saved — per element and per million-element tensor — appear alongside the rounding error so you can decide which pipelines can afford the cut.

  4. Precision Visualizer

    Enter any number and see the fp16 vs fp32 stored values, bit patterns, absolute and relative error, and a number-line showing the gap between representable Float16 values. Includes a tensor memory footprint slider.

  5. Audio sample compression

    One second of PCM audio (44100 samples) stored as Float32Array (172 KB) vs Float16Array (86 KB). Side-by-side waveforms show the compression is visually lossless. Choose from five signal types — sine, square, sawtooth, white noise, and speech envelope — and see max error, mean absolute error, estimated SNR, and a per-sample precision table.

why it shipped

Stage 3 TC39 proposal

references