Summary Statistics
3.6s
7B model restore
7.9s
13B model restore
391 MB/s
Peak decompress
100%
SHA-256 PASS rate
Detailed Results
| Model Size | Original | Compressed | Ratio | Compress | Decompress | Restore | Integrity |
|---|---|---|---|---|---|---|---|
| 1 GB (7B) | 1.0 GB | 875.1 MB | 1.17x | 208 MB/s | 283 MB/s | 3.615s | PASS |
| 3 GB (13B) | 3.0 GB | 2.4 GB | 1.24x | 237 MB/s | 391 MB/s | 7.861s | PASS |
| 7 GB (70B) | 7.0 GB | 5.7 GB | 1.23x | 289 MB/s | 378 MB/s | 18.979s | PASS |
Synthetic model weight data — float32 arrays with realistic distribution. Ratios reflect actual weight compressibility. Powered by HammerIO nvCOMP GPU LZ4.
Test Environment
- Hardware: NVIDIA Jetson AGX Orin 64 GB
- SDK: JetPack 6.x
- Compression: HammerIO nvCOMP GPU LZ4 (files >500 MB), CPU zstd (files <500 MB)
- Integrity: SHA-256 hash verification before and after every operation
- Data: Synthetic float32 model weight arrays with realistic distribution
Methodology
Each benchmark run consists of a full store-compress-verify-restore cycle. Timing begins at the start of the compress or decompress operation and ends when the operation completes and integrity is verified. All times are wall-clock. Three runs per model size; median reported.
Compression Strategy
- Files greater than 500 MB use nvCOMP GPU LZ4 via HammerIO for maximum throughput
- Files less than 500 MB use CPU zstd for efficient handling without GPU overhead
- SHA-256 hashes are computed before compression and verified after decompression
- All operations are logged to the ModelSafe manifest for audit trail