Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Boosting Scientific Error-Bounded Lossy Compression through Optimized Synergistic Lossy-Lossless Orchestration

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An open-source GPU compressor closes the gap with a proprietary one.

desk verdict A credible open-source GPU lossy compressor that beats the proprietary cuSZ-IB baseline in compression ratio, but the missing artifact link and in-sample pipeline selection need to be fixed before the headline numbers are fully convincing. read the letter →

arxiv 2507.11165 v2 pith:FIRMPNNW submitted 2025-07-15 cs.DC

Shixun Wu , Jinwen Pan , Jinyang Liu , Jiannan Tian , Ziwei Qiu , Jiajun Huang , Kai Zhao , Xin Liang
show 3 more authors
Sheng Di Zizhong Chen Franck Cappello
This is my paper
classification cs.DC
keywords error-boundedlossycompressionGPUinterpolationpredictionlosslessencodingquantizationcodereorderingscientificdatacuSZ-Hiratio
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

cuSZ-Hi is a GPU-based error-bounded lossy compressor that aims to close the compression-ratio gap between open-source GPU compressors and cuSZ-IB, the previously best GPU compressor, which depended on a proprietary NVIDIA encoding module. The paper argues that a carefully parallelized interpolation predictor, a reversible reordering of quantization codes by interpolation level, and two hand-tuned open-source lossless pipelines can together match or exceed cuSZ-IB's ratio while keeping throughput competitive. If right, this removes the need for a proprietary module to get high-ratio GPU lossy compression, making the technique usable in more environments. Reported gains are up to 249% compression ratio improvement at the same error bound and up to 215% at the same decompressed-data PSNR.

What carries the argument

The load-bearing mechanism is the synergy between an improved spline-interpolation predictor and the lossless stage. The predictor partitions data into $17^{3}$ blocks with anchor stride 16, runs hierarchical multi-dimensional interpolation (1D to 2D to 3D) with per-level auto-tuning among spline schemes, and then reorders the quantized integer errors by interpolation level so that codes with similar statistics are grouped. That reordered stream is fed to either HF-RRE4-TCMS8-RZE1 (Huffman plus run/zero-elimination and two's-complement-to-magnitude-sign transforms) or TCMS1-BIT1-RRE1 (magnitude-sign transform, bit shuffle, run elimination). The grouping is what lets simple open-source reducers approximate the redundancy removal that Bitcomp provided.

What would settle it

Run cuSZ-Hi-CR and cuSZ-IB on held-out scientific datasets from domains not used in the paper's pipeline selection, such as molecular dynamics trajectories or seismic gathers, at error bound 1e-3, and check whether cuSZ-Hi-CR still beats cuSZ-IB by the reported margin; a dataset where cuSZ-IB wins or the margin drops sharply would show the advantage is not universal.

Watch

Extended reading notes

Core claim

The central claim is that the two-stage recipe of lossy decomposition plus lossless encoding can be jointly optimized on GPUs so that an open-source pipeline beats the previously best GPU compressor, which relied on NVIDIA's proprietary Bitcomp. The paper shows this through cuSZ-Hi's two modes: a compression-ratio mode using Huffman encoding followed by the RRE4-TCMS8-RZE1 pipeline, and a throughput mode using TCMS1-BIT1-RRE1 without Huffman. On six benchmark datasets, cuSZ-Hi-CR reports the best compression ratio in most cases, with improvements over cuSZ-IB ranging beyond 200% in several error-bound settings, and cuSZ-Hi-TP improves throughput relative to cuSZ-I(B). The paper also claims this is the first GPU-based high-ratio, high-quality, and open-source scientific error-bounded lossy compressor.

Load-bearing premise

The two lossless pipelines were chosen by benchmarking on four datasets that also appear in the final evaluation, so the reported advantage over cuSZ-IB could shrink on data with different statistical structure.

Editorial extensions

If this is right

  • Open-source GPU lossy compressors can reach the compression ratio previously only achieved with a proprietary encoder, so high-ratio compression no longer requires vendor-specific libraries.
  • The same framework offers a compression-ratio mode and a throughput mode, so a single compressor can serve both storage-oriented and streaming-oriented scientific workloads.
  • The quantization-code reordering by interpolation level is a reusable technique for other interpolation-based compressors, since it exposes structure that entropy coding alone misses.
  • At large error bounds where the compression ratio exceeds roughly 300, the Huffman-tree overhead can make the throughput mode's ratio exceed the compression-ratio mode, showing the optimal pipeline depends on the operating point.
  • If the advantage holds across datasets, GPU-based workflows could compress exascale simulation output to a small fraction of its original size without sacrificing the error bound.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The lossless pipelines were selected by benchmarking on datasets that overlap the final evaluation set, so a fair reader should test on held-out data from other domains to see whether the 200%-plus gains generalize.
  • The reordering scheme's benefit likely depends on how strongly prediction error correlates with interpolation stride, so smoother fields should benefit more; this suggests a data-adaptive switch could be added.
  • A testable extension would be an online selector that picks the lossless pipeline per data block based on the measured quantization-code distribution, potentially improving both ratio and speed beyond the reported two modes.
  • If Bitcomp's advantage came largely from treating the same byte patterns, then this result implies that proprietary lossless modules may not be necessary for other interpolation-based compressors either.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents cuSZ-Hi, a GPU-based error-bounded lossy compressor with two operating modes: a compression-ratio-preferred mode (cuSZ-Hi-CR) and a throughput-preferred mode (cuSZ-Hi-TP). The design improves on the interpolation-based predictor of cuSZ-I through larger isotropic thread-block partitions, sparse anchor grids, multi-dimensional interpolation with auto-tuned configuration selection, and a quantization-code reordering step; it then couples this predictor with multi-stage lossless encoding pipelines (HF-RRE4-TCMS8-RZE1 and TCMS1-BIT1-RRE1) assembled from open-source components. The evaluation reports compression ratio, rate-distortion, throughput, and visual quality on six scientific datasets, comparing against cuSZ-L, cuSZ-I, cuSZ-IB, cuSZp2, cuZFP, and FZGPU, and claims up to 249% compression-ratio improvement under the same error bound and up to 215% improvement under the same decompression PSNR. An ablation study attributes the gains to the new data partition, code reordering, multi-dimensional interpolation/auto-tuning, and the lossless pipeline.

Significance. If the results hold, cuSZ-Hi would be a significant contribution: it is the first fully open-source GPU lossy compressor that approaches or exceeds the compression ratio of cuSZ-IB, which relies on NVIDIA's proprietary Bitcomp module. The paper's strengths include the open-source framing, a systematic benchmark of GPU lossless encoders (Figure 6), and a component-level ablation study (Table 5) that gives quantitative credit to each design choice. The throughput results on two GPU platforms also support the practical usability claim. However, the headline claims rest on the fairness of the cuSZ-IB baseline and on the risk that the lossless pipeline was selected using datasets that later appear in the final evaluation; both issues need to be resolved before the central comparison can be regarded as reliable.

major comments (3)
  1. [Table 1 vs. Table 4] There is an internal inconsistency in the NYX error-bound 1e-2 numbers. Table 1 reports that NVIDIA Bitcomp compresses the output of cuSZ-I (w/o Bitcomp) by 9.62x on the Nyx dataset at error bound 1e-2, while Table 4 reports cuSZ-I CR=29.5 and cuSZ-IB CR=249.0 on the same dataset and error bound. If Bitcomp is applied to the entire compressed stream of cuSZ-I, the expected cuSZ-IB CR is 29.5 × 9.62 ≈ 284, not 249.0. The authors should clarify whether Bitcomp is applied only to part of the intermediate stream (e.g., quantization codes excluding anchor points or Huffman metadata), whether Table 1 was measured on a different file or configuration, or why the multiplicative estimate does not apply. If the cuSZ-IB baseline is underconfigured relative to the Bitcomp capability measured in Table 1, the reported 236% advantage on this row is overstated, and the abstract's 'up to 249%' headline would need to be recomputed.
  2. [§5.2.2 and §6.2.1] The lossless pipeline selection appears to use the same datasets that later appear in the final evaluation. Section 5.2.2 and Figure 6 select the HF-RRE4-TCMS8-RZE1 pipeline by benchmarking on Hurricane, NYX, Miranda, and SCALE, and Table 4 then evaluates the full compressor on NYX and Miranda among other datasets. Since the pipeline choice is part of what drives the reported compression-ratio advantage over cuSZ-IB, this creates a selection-on-test-data risk: the chosen pipeline may be tuned to the benchmark family, and the reported gains may not generalize to other scientific datasets. I am not claiming deliberate overfitting, but the paper should either present a holdout evaluation (e.g., select the pipeline on one set of datasets and evaluate on a disjoint set) or report the results of all candidate pipelines on the final evaluation datasets so the reader can assess the sensitivity of the headline numbers.
  3. [§1, §6.1.2, and abstract] The abstract's claim of 'up to 249% compression ratio improvement' over 'existing state-of-the-art scientific lossy compressors' is broader than the evaluation. Section 1 states that cuSZ-I exhibits a much lower compression ratio than CPU-based SZ3 and QoZ, but Section 6.1.2 lists only GPU-based baselines (cuSZ variants, cuSZp2, cuZFP, FZGPU), and no CPU compressors are evaluated. The paper should either include CPU-based baselines such as SZ3 and QoZ in the comparison, or explicitly scope the 'up to 249%' and 'up to 215%' claims to GPU-based scientific lossy compressors. As written, the claims exceed the measured comparison set.
minor comments (5)
  1. [Table 5] There are typos in Table 5: 'finnally' should be 'finally', and the arrow notation '−−−→' with percentage labels is confusing; a standard incremental column layout would be clearer.
  2. [Eq. (3) and Eq. (4)] The reordering formula in Eq. (3) is difficult to read as printed; the indicator notation '1{x even}' and the prefix computation would benefit from a clean mathematical definition and a small 2D example to make the bijection explicit and verifiable.
  3. [§6.2.4] The throughput comparison in Figure 10 and the text reports speedup percentages, but there are no error bars or repeated-run information. Since GPU measurements can vary with clock throttling and initialization, the authors should state whether the reported numbers are single measurements or averages and, ideally, report variance.
  4. [Table 1] The capitalization in Table 1 is inconsistent with the rest of the paper: 'CUSZ-Hi-CR', 'CUSZ-L', and 'CUSZ-I' should be 'cuSZ-Hi-CR', 'cuSZ-L', and 'cuSZ-I' for consistency.
  5. [§6.2.2] The rate-distortion text contains unclear phrases such as '≈ 140%120% compression ratio improvement' and 'PSNR fixed around60/76'; these should be corrected to precise, unambiguous values.

Circularity Check

1 steps flagged · score 3.0 of 10

Lossless pipelines are selected on the same NYX/Miranda datasets that later appear in the headline CR table, making part of the measured advantage self-confirming; the 249% headline itself is from QMCPack, which was not in the selection benchmark.

  1. fitted input called prediction [Section 5.2.2, Figure 6; Section 6.2.1, Table 4]
    "For the LC framework, we perform preliminary experiments on several datasets and select 8 representative and adaptive pipelines (RRE1, RRE1-RRE2, etc., as shown in Figure 6) with 1/2/3/4 stages ... The benchmarking is conducted on 4 scientific datasets [51] ... According to the benchmarking results and analysis, we adopt the HF-RRE4-TCMS8-RZE1 lossless pipeline in cuSZ-Hi, forming its compression-ratio-preferred mode, cuSZ-Hi-CR."

    The lossless pipelines that define cuSZ-Hi-CR and cuSZ-Hi-TP were chosen by benchmarking compression ratio and throughput on Hurricane, NYX, Miranda, and SCALE (Figure 6). Table 4 then presents NYX and Miranda as evaluation cases supporting the claim that cuSZ-Hi outperforms baselines by up to 236%/219%. For these rows, the evaluated configuration is not an independent prediction but the argmax of the selection benchmark over the same datasets, so part of the reported advantage over cuSZ-IB is forced by the pipeline-selection criterion.

full rationale

Most of the paper's central claims are empirical measurements (Table 4, Figures 8 and 10) rather than derivations, so the compression-ratio numbers are not circular by construction. The main partial-circularity concern is the pipeline-selection overlap: the lossless encoders were benchmarked on NYX, Miranda, Hurricane, and SCALE, and NYX/Miranda reappear in the main CR evaluation, making those rows a selection-on-test rather than an independent test. This inflates, but does not by itself fabricate, the reported improvements. The headline 249% improvement is from QMCPack, a dataset not used in the lossless-pipeline selection, which limits the severity. Separately, the paper has an internal consistency issue that is a correctness/fairness concern rather than circularity: Table 1 reports Bitcomp CR 9.62 on cuSZ-I (w/o Bitcomp) output for NYX eb=1e-2, while Table 4 gives cuSZ-I CR 29.5 and cuSZ-IB CR 249.0; chaining the two stages implies roughly 284, suggesting the cuSZ-IB baseline may be underconfigured. The self-citations to cuSZ-I [36] and interpolation designs [33,34] are contextual and not load-bearing in the sense of a uniqueness theorem or an unverified premise that forces the results; the ablation study separately measures each component's contribution.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

cuSZ-Hi rests on a set of empirically chosen configuration parameters (block size, anchor stride, sample ratio, quantization width, pipeline stages, throughput threshold) and on domain assumptions about the compressibility of scientific data. No new physical or mathematical entities are introduced. The most load-bearing choices are the lossless pipeline stages, selected via benchmarking on datasets that overlap the main evaluation.

free parameters (7)
  • per-thread-block data block size = 17^3
    Chosen in preliminary experiments; authors state that larger blocks give limited CR gain and lower speed (Section 5.1.1).
  • anchor-point stride = 16
    Reduces anchor storage to 1/8 of cuSZ-I while preserving prediction quality (Section 5.1.1).
  • auto-tuning sample ratio = 0.2%
    Section 5.1.3 states the total sampled volume is 0.2%; no ablation of this ratio is provided.
  • quantization code width = uint8
    Section 5.2.1: codes resized to one-byte width; out-of-range outliers collected separately.
  • lossless pipeline stages = HF-RRE4-TCMS8-RZE1; TCMS1-BIT1-RRE1
    Selected via benchmarking on 4 datasets (Section 5.2.2, Figure 6); not derived from first principles.
  • Pareto throughput threshold = 25 GiB/s
    Excludes slower pipelines from the Pareto frontier (Section 5.2.2).
  • auto-tuned interpolation configurations = 1D/2D/3D splines, scheme set
    The set of tested splines and schemes is not enumerated; it is selected from prior work [33,34] and the auto-tuner chooses per level (Section 5.1.3).
assumptions (4)
  • domain assumption Spline interpolation provides accurate prediction of scientific data values
    Used throughout; the compressor relies on the spatial correlation of scientific data. Cited from [33,34,50] but not proven for the tested datasets.
  • ad hoc to paper Quantized prediction errors fit in uint8 with few outliers
    Assumed in Section 5.2.1; if the error histogram is not concentrated, the outlier list overhead grows and the compression ratio drops.
  • ad hoc to paper The reordering mapping in Eq.3 is a bijection
    The paper states the mapping groups codes by interpolation level, but the bijectivity is not proved, and the printed formula appears to contain typesetting errors.
  • domain assumption GPU hardware and CUDA behavior follow vendor documentation
    The implementation relies on CUDA, shared memory, and vendor primitives; this is a standard tooling assumption for GPU systems papers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting Scientific Error-Bounded Lossy Compression through Optimized Synergistic Lossy-Lossless Orchestration." pith.science (2026). https://pith.science/paper/FIRMPNNW

@misc{pith2026250711165,
  author       = {Pith},
  title        = {Pith review of: Boosting Scientific Error-Bounded Lossy Compression through Optimized Synergistic Lossy-Lossless Orchestration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FIRMPNNW}},
  note         = {Machine review of arXiv:2507.11165}
}
read the original abstract

As high-performance computing architectures evolve, more scientific computing workflows are being deployed on advanced computing platforms such as GPUs. These workflows can produce raw data at extremely high throughputs, requiring urgent high-ratio and low-latency error-bounded data compression solutions. In this paper, we propose cuSZ-Hi, an optimized high-ratio GPU-based scientific error-bounded lossy compressor with a flexible, domain-irrelevant, and fully open-source framework design. Our novel contributions are: 1) We maximally optimize the parallelized interpolation-based data prediction scheme on GPUs, enabling the full functionalities of interpolation-based scientific data prediction that are adaptive to diverse data characteristics; 2) We thoroughly explore and investigate lossless data encoding techniques, then craft and incorporate the best-fit lossless encoding pipelines for maximizing the compression ratio of cuSZ-Hi; 3) We systematically evaluate cuSZ-Hi on benchmarking datasets together with representative baselines. Compared to existing state-of-the-art scientific lossy compressors, with comparative or better throughput than existing high-ratio scientific error-bounded lossy compressors on GPUs, cuSZ-Hi can achieve up to 249% compression ratio improvement under the same error bound, and up to 215% compression ratio improvement under the same decompression data PSNR.

Figures

Figures reproduced from arXiv: 2507.11165 by the authors.

Figure 1
Figure 1. Design overview of cuSZ-I (interpolation-based sci [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Per-thread-block anchor and interpolation partition [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. GPU-based multi-dimensional interpolation (v.s. 1D [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Comparison between original and reordered quanti [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Lossless pipelines of cuSZ-Hi: HF-RRE4-TCMS8- [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: For each dataset, the upper rectangle plot displays [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visualization for quality evaluation. a similar compression speed and an acceptable ∼ 20% de￾compression time overhead compared to cuSZ-I(B) when achieving relatively higher compression ratios. On NVIDIA A100, the compression/decompression overhead of cuSZ-Hi￾CR compar…
Figure 10
Figure 10. Figure 10: Compression and decompression speed in GiB/s on NVIDIA Ada RTX 6000 and A100 with different error bounds. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 37 canonical work pages

  1. [1]

    Miranda application

    [n.d.]. Miranda application. https://wci.llnl.gov/simulation/ computer-codes/miranda

    work page

  2. [2]

    [n.d.]. nvCOMP. https://github.com/NVIDIA/nvcomp

    work page

  3. [3]

    NERSC Perlmutter - Lawrence Berkeley National Labratory

    2021. NERSC Perlmutter - Lawrence Berkeley National Labratory. https://docs.nersc.gov/systems/perlmutter/architecture/

    work page 2021

  4. [4]

    Frontier - Oak Ridge Leadership Computing Facility

    2022. Frontier - Oak Ridge Leadership Computing Facility. https: //www.olcf.ornl.gov/frontier/

    work page 2022

  5. [5]

    El Capitan: NNSA’s first exascale machine

    2023. El Capitan: NNSA’s first exascale machine. https://asc. llnl.gov/exascale/el-capitan

    work page 2023

  6. [6]

    Aurora - Argonne Leadership Computing Facility

    2024. Aurora - Argonne Leadership Computing Facility. https: //www.alcf.anl.gov/aurora

    work page 2024

  7. [7]

    Mark Ainsworth, Ozan Tugluk, Ben Whitney, and Scott Klasky

  8. [8]

    Noushin Azami, Alex Fallin, and Martin Burtscher. 2025. Efficient Lossless Compression of Scientific Floating-Point Data on CPUs and GPUs. InProceedings of the 30th ACM International Con- ference on Architectural Support for Programming Languages and Operating Systems, Volume 1(Rotterdam, Netherlands) (ASPLOS ’25). Association for Computing Machinery, New...

    arXiv 2025

Show all 57 references
  1. [9]

    Rafael Ballester-Ripoll, Peter Lindstrom, and Renato Pajarola

  2. [10]

    David Carrasco-Busturia, Emiliano Ippoliti, Simone Meloni, Ur- sula Rothlisberger, and Jógvan Magnus Haugaard Olsen. 2024. Multiscale biomolecular simulations in the exascale era.Cur- rent Opinion in Structural Biology86 (2024), 102821. https: //doi.org/10.1016/j.sbi.2024.102821

    2024

  3. [11]

    Xinyu Chen, Jiannan Tian, Ian Beaver, Cynthia Freeman, Yan Yan, Jianguo Wang, and Dingwen Tao. 2024. FCBench: Cross- Domain Benchmarking of Lossless Compression for Floating-Point Data. Proceedings of the VLDB Endowment17, 6 (2024), 1418– 1431

    2024

  4. [12]

    Yann Collet. 2015. Zstandard – Real-time data compression algorithm. http://facebook.github.io/zstd/ (2015)

    2015

  5. [13]

    cuZFP. 2020. https://github.com/LLNL/zfp/tree/develop/src/ cuda_zfp. Online

    2020

  6. [14]

    L Peter Deutsch. 1996. GZIP file format specification version 4.3

    1996

  7. [15]

    1996.DEFLATE compressed data format speci- fication version 1.3

    Peter Deutsch. 1996.DEFLATE compressed data format speci- fication version 1.3. Technical Report

    1996

  8. [16]

    Qian Gong, Jieyang Chen, Ben Whitney, Xin Liang, Viktor Reshniak, Tania Banerjee, Jaemoon Lee, Anand Rangarajan, Lipeng Wan, Nicolas Vidal, Qing Liu, Ana Gainaru, Norbert Podhorszki, Richard Archibald, Sanjay Ranka, and Scott Klasky

  9. [17]

    Lucas Hayne, John Clyne, and Shaomeng Li. 2021. Using Neural Networks for Two Dimensional Scientific Data Compression. In 2021 IEEE International Conference on Big Data (Big Data). IEEE, 2956–2965

    2021

  10. [19]

    Langwen Huang and Torsten Hoefler. 2022. Compressing multidi- mensional weather and climate data into neural networks.arXiv preprint arXiv:2210.12538 (2022)

    2022 arXiv

  11. [20]

    Yafan Huang, Sheng Di, Guanpeng Li, and Franck Cappello. 2024. cuSZp2: A GPU Lossy Compressor with Extreme Throughput and Optimized Compression Ratio. InSC24: International Con- ference for High Performance Computing, Networking, Storage and Analysis. IEEE, 1–18

    2024

  12. [21]

    Yafan Huang, Sheng Di, Xiaodong Yu, Guanpeng Li, and Franck Cappello. 2023. cuSZp: An Ultra-fast GPU Error-bounded Lossy Compression Framework with Optimized End-to-End Perfor- mance. InProceedings of the International Conference for High Performance Computing, Networking, Sto...

    2023

  13. [22]

    J. E. Kay and et al. 2015. The Community Earth System Model (CESM) large ensemble project: A community resource for study- ing climate change in the presence of internal climate variability. Bulletin of the American Meteorological Society96, 8 (2015), 1333–1349

    2015

  14. [23]

    Suha Kayum et al. 2020. GeoDRIVE – a high performance computing flexible platform for seismic applications.First Break 38, 2 (2020), 97–100

    2020

  15. [24]

    Jeongnim Kim et al. 2018. QMCPACK: an open source ab initio quantum Monte Carlo package for the electronic structure of atoms, molecules and solids. Journal of Physics: Condensed Matter 30, 19 (2018), 195901. https://doi.org/10.1088/1361- 648x/aab9c3

    2018 doi

  16. [25]

    Fabian Knorr, Peter Thoman, and Thomas Fahringer. 2021. ndzip- gpu: efficient lossless compression of scientific floating-point data on GPUs. InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analy- sis. 1–14

    2021

  17. [26]

    Shaomeng Li, Peter Lindstrom, and John Clyne. 2023. Lossy scientific data compression with SPERR. In2023 IEEE Interna- tional Parallel and Distributed Processing Symposium (IPDPS). IEEE, 1007–1017

    2023

  18. [27]

    Xiao Li, Jaemoon Lee, Anand Rangarajan, and Sanjay Ranka

  19. [28]

    Yi Li, Eric Perlman, Minping Wan, Yunke Yang, Charles Mene- veau, Randal Burns, Shiyi Chen, Alexander Szalay, and Gregory Eyink. 2008. A public turbulence database cluster and appli- cations to study Lagrangian evolution of velocity increments in turbulence. Journal of Turbule...

    2008

  20. [29]

    Gok, Jiannan Tian, Junjing Deng, Jon C

    Xin Liang, Kai Zhao, Sheng Di, Sihuan Li, Robert Underwood, Ali M. Gok, Jiannan Tian, Junjing Deng, Jon C. Calhoun, Ding- wen Tao, Zizhong Chen, and Franck Cappello. 2023. SZ3: A Mod- ular Framework for Composing Prediction-Based Error-Bounded Lossy Compressors.IEEE Transactio...

    2023

  21. [30]

    Peter Lindstrom. 2014. Fixed-rate compressed floating-point ar- rays. IEEE transactions on visualization and computer graphics 20, 12 (2014), 2674–2683

    2014

  22. [31]

    Jinyang Liu, Sheng Di, Sian Jin, Kai Zhao, Xin Liang, Zizhong Chen, and Franck Cappello. 2023. Scientific Error-bounded Lossy Compression with Super-resolution Neural Networks. In2023 IEEE International Conference on Big Data (BigData). IEEE, 229–236

    2023

  23. [32]

    Jinyang Liu, Sheng Di, Kai Zhao, Sian Jin, Dingwen Tao, Xin Liang, Zizhong Chen, and Franck Cappello. 2021. Exploring Autoencoder-based Error-bounded Compression for Scientific Data. In2021 IEEE International Conference on Cluster Com- puting (CLUSTER). IEEE, 294–306

    2021

  24. [33]

    Jinyang Liu, Sheng Di, Kai Zhao, Xin Liang, Zizhong Chen, and Franck Cappello. 2022. Dynamic quality metric oriented error bounded lossy compression for scientific datasets. In2022 SC22: International Conference for High Performance Computing, Networking, Storage and Analysis ...

    2022

  25. [34]

    Jinyang Liu, Sheng Di, Kai Zhao, Xin Liang, Sian Jin, Zizhe Jian, Jiajun Huang, Shixun Wu, Zizhong Chen, and Franck Cap- pello. 2024. High-performance effective scientific error-bounded lossy compression with auto-tuned multi-component interpolation. Proceedings of the ACM on ...

    2024

  26. [35]

    Jinyang Liu, Sihuan Li, Sheng Di, Xin Liang, Kai Zhao, Dingwen Tao, Zizhong Chen, and Franck Cappello. 2021. Improving Lossy Compression for SZ by Exploring the Best-Fit Lossless Com- pression Techniques. In2021 IEEE International Conference on Big Data (Big Data). 2986–2991. ...

    2021

  27. [36]

    Jinyang Liu, Jiannan Tian, Shixun Wu, Sheng Di, Boyuan Zhang, Robert Underwood, Yafan Huang, Jiajun Huang, Kai Zhao, Guan- peng Li, Dingwen Tao, Zizhong Chen, and Franck Cappello. 2024. CUSZ-i: High-Ratio Scientific Lossy Compression on GPUs with Optimized Multi-Level Interpol...

    2024 arXiv

  28. [37]

    Youyuan Liu, Wenqi Jia, Taolue Yang, Miao Yin, and Sian Jin

  29. [38]

    Yuzhe Lu, Kairong Jiang, Joshua A Levine, and Matthew Berger

  30. [39]

    Nyx simulation. 2019. https://amrex-astro.github.io/Nyx. On- line

    2019

  31. [40]

    William A Pearlman, Asad Islam, Nithin Nagaraj, and Amir Said. 2004. Efficient, low-complexity image coding with a set- partitioning embedded block coder.IEEE transactions on cir- cuits and systems for video technology14, 11 (2004), 1219–1235

    2004

  32. [41]

    Cody Rivera, Sheng Di, Jiannan Tian, Xiaodong Yu, Dingwen Tao, and Franck Cappello. 2022. Optimizing Huffman Decoding for Error-Bounded Lossy Compression on GPUs. In2022 IEEE International Parallel and Distributed Processing Symposium (IPDPS). 717–727. https://doi.org/10.1109/...

    2022

  33. [42]

    InSC24-W: Workshops of the International Conference for High Performance Com- puting, Networking, Storage and Analysis

    Enhancing Lossy Compression Through Cross-Field In- formation for Scientific Applications. InSC24-W: Workshops of the International Conference for High Performance Com- puting, Networking, Storage and Analysis. 300–308. https: //doi.org/10.1109/SCW63240.2024.00046

  34. [43]

    Dingwen Tao, Sheng Di, Hanqi Guo, Zizhong Chen, and Franck Cappello. 2019. Z-checker: A framework for assessing lossy com- pression of scientific data.The International Journal of High Performance Computing Applications 33, 2 (2019), 285–303. https://doi.org/10.1177/1094342017737147

    2019 doi

  35. [44]

    Jiannan Tian, Sheng Di, Xiaodong Yu, Cody Rivera, Kai Zhao, Sian Jin, Yunhe Feng, Xin Liang, Dingwen Tao, and Franck Cappello. 2021. cuSZ (x): Optimizing Error-Bounded Lossy Com- pression for Scientific Data on GPUs.CoRR (2021)

    2021

  36. [45]

    Jiannan Tian, Sheng Di, Kai Zhao, Cody Rivera, Megan Hick- man Fulp, Robert Underwood, Sian Jin, Xin Liang, Jon Cal- houn, Dingwen Tao, and Franck Cappello. 2020. cuSZ: An Ef- ficient GPU-Based Error-Bounded Lossy Compression Frame- work for Scientific Data. In Proceedings of ...

    2020

  37. [46]

    Yeung, Kiran Ravikumar, Stephen Nichols, and Rohini Uma- Vaideswaran

    P.K. Yeung, Kiran Ravikumar, Stephen Nichols, and Rohini Uma- Vaideswaran. 2025. GPU-enabled extreme-scale turbulence simu- lations: Fourier pseudo-spectral algorithms at the exascale using OpenMP offloading. Computer Physics Communications306 (2025), 109364. https://doi.org/1...

    2025

  38. [47]

    Xiaodong Yu, Sheng Di, Kai Zhao, jiannan Tian, Dingwen Tao, Xin Liang, and Franck Cappello. 2022. SZx: an Ultra-fast Error- bounded Lossy Compressor for Scientific Datasets.arXiv preprint arXiv:2201.13020 (2022)

    2022 arXiv

  39. [48]

    Dingwen Tao, Sheng Di, Zizhong Chen, and Franck Cappello. 2017. Significantly improving lossy compression for scientific data sets based on multidimensional prediction and error-controlled quan- tization. In2017 IEEE International Parallel and Distributed Processing Symposium....

    2017

  40. [49]

    Boyuan Zhang, Jiannan Tian, Sheng Di, Xiaodong Yu, Martin Swany, Dingwen Tao, and Franck Cappello. 2023. GPULZ: Opti- mizing LZSS Lossless Compression for Multi-byte Data on Modern 13 Shixun Wu, Jinwen Pan, Jinyang Liu, Jiannan Tian, et al. GPUs. InProceedings of the 37th Inte...

    2023

  41. [50]

    Tonel- lot, Zizhong Chen, and Franck Cappello

    Kai Zhao, Sheng Di, Maxim Dmitriev, Thierry-Laurent D. Tonel- lot, Zizhong Chen, and Franck Cappello. 2021. Optimizing Error-Bounded Lossy Compression for Scientific Data by Dy- namic Spline Interpolation. In2021 IEEE 37th International Conference on Data Engineering (ICDE). 1...

    2021

  42. [51]

    Kai Zhao, Sheng Di, Xin Lian, Sihuan Li, Dingwen Tao, Julie Bessac, Zizhong Chen, and Franck Cappello. 2020. SDRBench: Scientific Data Reduction Benchmark for Lossy Compressors. In 2020 IEEE International Conference on Big Data (Big Data). 2716–2724

    2020

  43. [52]

    Kai Zhao, Sheng Di, Xin Liang, Sihuan Li, Dingwen Tao, Zizhong Chen, and Franck Cappello. 2020. Significantly Improving Lossy Compression for HPC Datasets with Second-Order Prediction and Parameter Optimization. InProceedings of the 29th Inter- national Symposium on High-Perfo...

    2020

  44. [54]

    Boyuan Zhang, Jiannan Tian, Sheng Di, Xiaodong Yu, Yunhe Feng, Xin Liang, Dingwen Tao, and Franck Cappello. 2023. FZ- GPU: A Fast and High-Ratio Lossy Compressor for Scientific Com- puting Applications on GPUs.arXiv preprint arXiv:2304.12557 (2023)

    2023 arXiv

  45. [2018]

    Multilevel techniques for compression and reduction of scientific data—the univariate case.Computing and Visualization in Science19, 5 (2018), 65–76

    2018

  46. [2019]

    IEEE transactions on visualization and computer graphics 26, 9 (2019), 2891–2903

    TTHRESH: Tensor compression for multidimensional visual data. IEEE transactions on visualization and computer graphics 26, 9 (2019), 2891–2903

    2019

  47. [2021]

    In Computer Graphics Forum, Vol

    Compressive neural representations of volumetric scalar fields. In Computer Graphics Forum, Vol. 40. Wiley Online Library, 135–146

  48. [2023]

    https://doi.org/10.1016/j.softx.2023.101590

    MGARD: A multigrid framework for high-performance, error-controlled data compression and refactoring.SoftwareX 24 (2023), 101590. https://doi.org/10.1016/j.softx.2023.101590

    2023

  49. [2024]

    In2024 IEEE In- ternational Conference on Big Data (BigData)

    Attention Based Machine Learning Methods for Data Reduction with Guaranteed Error Bounds. In2024 IEEE In- ternational Conference on Big Data (BigData). 1039–1048. https://doi.org/10.1109/BigData62323.2024.10825655

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.