GPU-accelerated compression for edge AI. GPU where it matters. CPU where it doesn't.
Created by ResilientMind AI | ResilientMindai.com | Joseph C McGinty Jr
HammerIO is a GPU-accelerated compression toolkit that automatically routes files to the fastest available hardware. Large files go to nvCOMP GPU LZ4 (6+ GB/s decompression). On macOS, it uses Apple LZFSE via the Accelerate framework. Everything else goes to CPU zstd with parallel threading. No flags, no configuration.
pip install hammerioGPU acceleration (NVIDIA Jetson or desktop GPU):
pip install hammerio[gpu]Jetson with jtop monitoring:
pip install hammerio[jetson]From source:
git clone https://github.com/Subzero121800/HammerIO.git
cd HammerIO && pip install -e .Requires Python 3.10+. GPU features require CUDA 12.x and NVIDIA nvCOMP.
After installing with pip, run these optional setup commands:
# Accept the license agreement (shown automatically on first run)
hammer --accept-license
# Install right-click compress/decompress in your file manager
hammer install-desktop
# Launch the web dashboard (GPU monitoring, file browser, compression UI)
hammer webui# Compress anything -- routing is automatic
hammer compress data.csv
hammer compress ./dataset/ --quality fast
hammer compress archive.tar --algo lz4
# Decompress
hammer decompress data.csv.zst
hammer decompress archive.tar.lz4
# Batch compress a directory
hammer batch ./logs/ --workers 8
# See what HammerIO would do
hammer info --routes ./my_file.csv
# Hardware profile
hammer info --hardware
# Install right-click context menu (Nautilus, Nemo, Thunar)
hammer install-desktopPython API:
import hammerio
router = hammerio.JobRouter(quality="fast")
job = router.route("dataset.csv")
result = router.execute(job)
print(f"{result.compression_ratio:.1f}x via {result.processor_used}")Measured on Jetson AGX Orin 64GB, JetPack 6.2.2, CUDA 12.6, MAXN mode. Realistic mixed data (25% CSV, 25% patterns, 25% semi-random, 25% true entropy).
| Method | Compress | Decompress | Ratio | Roundtrip |
|---|---|---|---|---|
| CPU zstd-1 | 1,094 MB/s | 733 MB/s | 2.00x | 24.1s |
| CPU zstd-3 | 1,014 MB/s | 741 MB/s | 2.00x | 24.0s |
| GPU nvCOMP LZ4 | 517 MB/s | 4,258 MB/s | 1.98x | 22.2s |
GPU wins the roundtrip: 1.1x faster than CPU best. GPU decompression is 5.7x faster than CPU.
| Method | Compress | Decompress | Ratio |
|---|---|---|---|
| CPU zstd-1 | 1,747 MB/s | 2,001 MB/s | 2.00x |
| CPU zstd-3 | 1,615 MB/s | 2,088 MB/s | 2.00x |
| GPU nvCOMP LZ4 | 705 MB/s | 8,537 MB/s | 1.98x |
| GPU nvCOMP Snappy | 1,615 MB/s | 5,756 MB/s | 1.68x |
| Pattern | IOPS | Bandwidth | Latency p99 |
|---|---|---|---|
| Random Read | 109,299 | 6,831 MB/s | 0.02 ms |
| Random Write | 29,405 | 1,838 MB/s | 0.05 ms |
| Mixed 70/30 R/W | 57,146 | 3,572 MB/s | 0.05 ms |
| Size | CPU zstd-3 | GPU LZ4 | GPU Wins? |
|---|---|---|---|
| 1 MB | 260 MB/s | 29 MB/s | No |
| 10 MB | 341 MB/s | 378 MB/s | Yes |
| 100 MB | 678 MB/s | 693 MB/s | Yes |
| 1 GB | 380 MB/s | 464 MB/s | Yes |
GPU crossover at ~10 MB. Below that, kernel launch overhead dominates.
Run benchmarks yourself:
hammer benchmark # Standard (500MB)
hammer benchmark --quick # Quick (100MB)
hammer benchmark --10gb # Stress test (10GB)
hammer benchmark --type memory # In-memory only
hammer benchmark --type random-io # Random I/O patterns
hammer benchmark --type scale # Scalability sweepHammerIO profiles every file and routes to the optimal compressor:
$ hammer info --hardware
Routing Profile:
Large Files -> nvCOMP LZ4 (GPU) # Files > 500MB
Datasets -> nvCOMP LZ4 (GPU) # CSV, Parquet, NPY, etc.
General -> zstd parallel (CPU) # Default path
Archives -> passthrough # Already compressed
Text Logs -> zstd (CPU, high ratio) # Best ratio
On macOS Apple Silicon: General -> Apple LZFSE (Accelerate) -- hardware-optimized, faster than zstd on M-series.
If the GPU path fails (OOM, driver issue), HammerIO falls back to CPU and logs why.
Drop-folder automation for edge pipelines:
hammer watch --watch-root ./pipeline --threshold-mb 500
# ./pipeline/
# compress/ <- drop files here
# decompress/ <- drop .zst/.lz4 files here
# compressed/ <- output appears here
# decompressed/ <- output appears here
# processed/ <- originals moved here after successGPU routing kicks in for files above the threshold. CPU handles the rest.
hammer webui # http://localhost:5000Real-time monitoring: GPU/CPU utilization, thermal zones, power rails, per-core CPU bars, compression job history, file browser, quick compress/decompress, and a built-in CLI console.
| Device | Compression Engine | Status |
|---|---|---|
| Jetson AGX Orin | nvCOMP GPU LZ4 | Primary target, fully tested |
| Jetson Orin NX / Nano | nvCOMP GPU LZ4 | Supported |
| RTX 3000 / 4000 / 5000 | nvCOMP GPU LZ4 | Supported |
| Any CUDA 12.x GPU | nvCOMP GPU LZ4 | Supported |
| macOS Apple Silicon | Apple LZFSE (Accelerate) | M1/M2/M3/M4 |
| CPU-only (no GPU) | zstd / gzip / lz4 | Universal fallback |
| Command | Description |
|---|---|
hammer compress |
Compress file or directory (auto GPU/CPU routing) |
hammer decompress |
Decompress .zst, .lz4, .lzfse, .gz, .bz2 files |
hammer batch |
Batch compress a directory with parallel workers |
hammer watch |
Watch folders and auto-process dropped files |
hammer benchmark |
Run GPU vs CPU benchmark suite |
hammer install-desktop |
Install right-click compress/decompress in file manager |
hammer info |
Hardware profile, routing decisions, telemetry |
hammer config |
Show, save, or generate configuration |
hammer monitor |
Live terminal telemetry (jtop-style) |
hammer webui |
Launch web dashboard |
hammer version |
Version and system info |
Copyright 2026 ResilientMind AI | ResilientMindai.com | Joseph C McGinty Jr
Proprietary License — All Rights Reserved
Free for personal, non-commercial evaluation only.
Commercial licenses available for redistribution, SaaS, and OEM embedding.
| License | Price |
|---|---|
| Individual commercial | $199 one-time |
| Organization | $999/year |
| OEM / Embedded | Custom |
See COMMERCIAL_LICENSE.md for details.
Contact: Joseph@ResilientMindAI.com | resilientmindai.com
ResilientMind AI | resilientmindai.com | Joseph C McGinty Jr