Fuga
Fuga stores identical files once. It splits the remaining files at content-defined boundaries so that repeated fragments can be shared even when bytes are inserted elsewhere in a file. It compresses the unique data in blocks. A catalogue lets you restore a selected file or folder.
Reclaim disk space
Verify the archive against the source byte for byte. The standard .fg workflow can prepare a list of originals to remove and checks them again after you approve that exact list. It supports split volumes, interrupted writes that can resume, and optional XChaCha20-Poly1305 encryption for file contents and catalogue.
CPU, GPU and NPU
Fuga is a standalone CLI for macOS, Linux and Windows. In the 2.4 source, CPU is the default. On compatible hardware, use fuga create source backup --backend gpu or --backend npu. Framed .fg archives can be restored on CPU. This mode uses temporary disk space and currently supports single-volume archives without encryption or source release.
Metal runs match search and Huffman bitstream operations on supported Macs. The NPU computes prediction scores through Core ML or a configured USB Coral on Linux; rank coding and feedback still run on CPU. Reusable dictionary packs represent repeated fragments and can be embedded or stored separately. External packs must be kept for restoration.
Archiver comparison
Compare only rows from the same test. x86 used 533.8 MB of project files on a 2-vCPU Ubuntu VM. Mac A used 629 text, code, binary and media files (26.65 MB); Mac B and Pi B exclude media. Mac rates are from one run on a shared Mac; Pi rates are medians of three repetitions. Ratios include required dictionaries and models where used.
| Format / setting | Test | Ratio | Pack MiB/s | Extract MiB/s |
|---|---|---|---|---|
| tar.gz -6 (gzip) | x86 | 2.61× | 25.75 | 109.39 |
| tar.bz2 -9 (bzip2) | x86 | 2.84× | 7.25 | 12.59 |
| zip -6 (Info-ZIP) | x86 | 2.53× | 26.61 | 55.79 |
| rar -m5 solid | x86 | 6.61× | 6.58 | 53.79 |
| 7z LZMA2 -mx9 | x86 | 7.95× | 2.96 | 58.08 |
| tar.xz -9e (xz) | x86 | 7.60× | 1.53 | 37.19 |
| tar.zst -3 (zstd) | x86 | 2.98× | 191.78 | 62.27 |
| tar.zst -19 long (zstd) | x86 | 7.37× | 3.43 | 201.23 |
| Fuga fast | x86 | 7.49× | 19.07 | 57.51 |
| Fuga max | x86 | 8.05× | 3.81 | 62.94 |
| tar.zst -7 | Mac A | 2.012× | 7.844 | 12.515 |
| Fuga CPU · local Huffman | Mac A | 1.461× | 1.796 | 7.073 |
| Fuga GPU · Metal Huffman | Mac A | 1.461× | 4.219 | 5.744 |
| Fuga CPU · rank Huffman | Mac B | 1.245× | 4.382 | 5.971 |
| Fuga NPU · Core ML Huffman | Mac B | 1.245× | 2.583 | 3.905 |
| tar.zst -3 | Pi B | 2.261× | 53.47 | 189.17 |
| Fuga CPU · rank zstd-3 | Pi B | 2.054× | 21.69 | 65.24 |
| Fuga NPU · USB Coral, zstd-3 | Pi B | 1.898× | 21.09 | — |
On the Mac, Metal packed 2.35× as fast as the matched CPU Huffman run, extracted more slowly and produced identical bytes. Core ML was slower both ways; its plan preferred the Neural Engine, with no hardware-counter confirmation. Physical USB Coral packed slightly more slowly than its matched CPU run. The measured Coral archives used the CPU for restoration.
ZIP is convenient for exchange. On the x86 corpus, tar.zst -3 packed fastest and Fuga max made the smallest archive of the rows shown; 7z was close. RAR, tar.gz, tar.bz2 and tar.xz are shown for comparison. Apple Archive was tested separately on macOS.
All software