FMT Lossless Codec Beats NVIDIA HTJ2K on Size and Speed
TL;DR
- FMT decodes images of any size in a fixed number of sequential steps, using a coarse-to-fine tree paired with parallel rANS entropy-decoding lanes.
- Without any neural network, the codec beats NVIDIA's HTJ2K on both file size and decode speed, the single-author paper reports.
- With a large network, FMT decodes Kodak photos 9 to 61 times faster than other learned codecs running on the same A100 GPU, at the same size or smaller.
FMT, a lossless image codec described in a single-author preprint posted on November 18 by Ammar Ali, decodes an image of any size in a fixed number of sequential steps. The author credits that property to designing the format's decoding order and its entropy coder together, so neither the network nor the arithmetic decoder has to wait on dependencies that grow with resolution.
The abstract puts the comparisons bluntly. "Without a network, FMT beats NVIDIA's HTJ2K on size and speed," Ali writes, and "with a large one, it decodes Kodak photos 9 to 61X faster than learned codecs we ran on the same A100, at the same size or smaller." He also claims "no codec we ran on the same GPU beats any of them on both size and speed."
The underlying argument is about waiting rather than compute. "Much of this time is spent waiting, not computing: the network waits for the values it conditions on, and the arithmetic decoder reads them one at a time," the paper says. FMT stores an image as a coarse-to-fine tree whose steps hold independent values, which a GPU predicts at once and decodes through parallel rANS lanes that carry their states from step to step.
It lands in a steady stream of open-source codec and systems papers we have tracked this quarter, though the comparisons here are limited to the sets and A100 hardware the author actually ran.
Originally reported by huggingface.co
Read the original article →Original headline: FMT Lossless Image Codec Decodes in at Most 72 Parallel Steps, Beats NVIDIA HTJ2K on Size and Speed