FSZ GPU compressor crushes ratios 10.9x faster without throughput loss
New GPU compressor achieves 10x better compression at record 676 GB/s throughput.
Traditional GPU lossy compressors sacrifice compression ratio for speed, using fixed first-order predictors on independent blocks. FSZ, developed by Jiajun Huang and accepted at SC26, breaks this trade-off with three algorithmic innovations within a single CUDA kernel. Cross-block prediction carries Lorenzo prediction state across 256-element tile boundaries, eliminating 7 out of 8 boundary residuals that inflate encoding rates. Per-tile adaptive multi-order prediction selects the best strategy per tile from first-order, second-order, and centering variants. Finally, a single-pass four-way evaluation leverages finite differences to evaluate all variants from a single data read, enabling richer prediction without extra bandwidth.
On NVIDIA GH200 with eight real-world datasets, FSZ outperforms cuSZp-P by up to 10.95x and the state-of-the-art cuSZp-O by up to 2.92x in compression ratio. Remarkably, these gains come with no throughput penalty: FSZ achieves the highest average throughput among all evaluated compressors—676 GB/s for compression and 785 GB/s for decompression. This makes FSZ ideal for exascale simulations, climate modeling, and other data-intensive scientific workloads where both speed and storage efficiency are critical.
- Cross-block prediction eliminates 7 out of 8 boundary residuals by carrying Lorenzo prediction state across 256-element tile boundaries.
- Per-tile adaptive multi-order prediction selects the best strategy from first-order, second-order, and centering variants per tile.
- On NVIDIA GH200, FSZ achieves 676 GB/s compression throughput and 785 GB/s decompression, highest among evaluated compressors.
Why It Matters
Enables scientists to store and transfer massive simulation datasets with unprecedented speed and compression efficiency.