Pith. sign in

REVIEW 3 major objections 5 minor 36 references

FSZ: Breaking the Prediction-Throughput Trade-off in GPU Lossy Compression

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read FSZ claims a single CUDA kernel can deliver both the highest compression ratio and the highest throughput among GPU lossy compressors, by exploiting the idle arithmetic in memory-bound passes.

desk verdict A genuine systems contribution with a clever single-pass trick and a broad evaluation; the headline trade-off claim is plausible but the 'free arithmetic' mechanism isn't isolated — worth refereeing. read the letter →

arxiv 2607.15413 v1 pith:63FOJIPL submitted 2026-07-16 cs.DC cs.PF

Jiajun Huang This is my paper
classification cs.DCcs.PF
keywords GPUlossycompressionerror-boundedLorenzopredictionsingle-kernelCUDAthroughputratioscientificdataadaptive
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

The paper aims to overturn the assumption that richer prediction in a fast GPU lossy compressor must cost throughput. It proposes FSZ, a single-CUDA-kernel error-bounded compressor that carries prediction state across block boundaries, adaptively picks among four prediction variants per tile, and evaluates all four in a single data read by exploiting the fact that constant offsets cancel in finite differences. The claim is that the first pass of two-pass variable-length encoding is memory-bandwidth-bound, leaving arithmetic headroom, so the richer prediction runs at no extra bandwidth cost. If true, GPU-accelerated scientific and AI pipelines can get meaningfully better compression ratios without slowing down inline compression, which matters at the hundreds of GB/s rates these workloads need.

What carries the argument

The key object is the single-pass four-way evaluation, built on the finite-difference cancellation identity: for k-th order Lorenzo (finite-difference) prediction, Δ^k(q_i − μ) = Δ^k q_i for a constant offset μ, so centering changes only the first one or two residuals of a tile. That lets a thread computing LZ1 and LZ2 residuals in one loop also derive the centering costs with a trivial boundary adjustment, without additional global reads. The other carrier is cross-block prediction state, maintained by having one thread process all 8 blocks of a 256-element tile sequentially so state lives in registers — a zero-sync resolution of the sequential dependency. Together they fill the arithmetic

What would settle it

A register-usage/occupancy measurement of the FSZ kernel: if the added LZ2 state and centering accumulator reduce achieved occupancy below a baseline first-order kernel, or if running FSZ on small (e.g., sub-1MB) latency-bound inputs shows a throughput gap that grows with prediction richness, the headroom assumption would be falsified. Concretely, compare FSZ's per-kernel time against a variant that evaluates only LZ1 with independent blocks on arrays of 1KB, 1MB, and 1GB; the richness penalty should appear in the small-array regime if the claim is universal.

Watch

Extended reading notes

Core claim

The central claim is that the prediction-throughput trade-off in pure-GPU error-bounded lossy compression is not fundamental. FSZ's three innovations work together: cross-block prediction state reduces boundary residuals from 8 per 256-element tile to 1; per-tile adaptive selection chooses between first- and second-order Lorenzo prediction with or without centering; and single-pass four-way evaluation computes all four variant costs from a single read of the data, because a constant offset cancels exactly in first- and second-order finite differences for all but the first one or two elements. The author argues the rate pass is memory-bound rather than compute-bound, so the extra arithmetic h

Load-bearing premise

The whole 'no throughput penalty' result rests on the claim that the first pass is memory-bandwidth-bound, not compute-bound, so the extra prediction arithmetic hides in memory-stall cycles without hurting occupancy or adding time — if that balance shifts (small arrays, different hardware, tighter register budgets), the trade-off reappears.

Editorial extensions

If this is right

  • FSZ improves compression ratio by up to 10.95x over cuSZp-P and 2.92x over cuSZp-O across 8 scientific datasets, with the largest gains on fields with linear/polynomial trends or constant offsets.
  • The throughput advantage holds on both GH200 and A100, and the speedup over cuSZp is larger on A100 because the higher ratio writes fewer bytes, amplifying the benefit on lower-bandwidth memory.
  • Rate-0 blocks can be skipped and their predictor state advanced analytically, so blocks of all-zero residuals cost nothing in either compressed size or decode work.
  • All four prediction variants preserve the same pointwise error bound because quantization is the only lossy step; centering is an exact invertible integer transform.
  • Adaptive centering is enabled only when its 4-byte mean cost is outweighed by savings, so the per-tile selection is purely a compression-ratio decision.

Reading between the lines

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

  • The same cancellation trick should generalize to higher-order finite-difference predictors and longer prediction chains: if the arithmetic headroom holds, incorporating third-order terms or longer tiles would extend the ratio gains without extra bandwidth — but register pressure and growing boundary-adjustment arithmetic would eventually hit a compute or occupancy wall.
  • The headroom argument is hardware-specific: on a GPU with a very different ALU-to-memory ratio, or on latency-bound workloads (small arrays, launch-overhead-dominated kernels), the hidden work may no longer hide. A portable extension would need an adaptive 'richness' knob, like turning off LZ2 or widening tiles on smaller inputs.
  • The cross-block state trick could transfer to other block-based GPU algorithms with sequential dependencies (e.g., finite-difference stencil computation, prefix-sum-like scans) where a single thread processing a contiguous run in registers removes synchronization costs.
  • Testing FSZ on arrays whose values follow a known polynomial (e.g., an exact quadratic) would cleanly separate the contribution of LZ2 from cross-block state; the paper's ablation on real fields already suggests LZ2 dominates at tight bounds.
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. FSZ is a single-CUDA-kernel GPU lossy compressor that combines cross-block Lorenzo prediction state over 256-element tiles, per-tile adaptive selection among four prediction variants (LZ1/LZ2 with/without centering), and a single-pass four-way cost evaluation based on exact cancellation of constant offsets in finite differences. The paper reports compression-ratio gains of up to 2.92x over cuSZp-O and 10.95x over cuSZp-P on eight SDRBench/Open-SciVis datasets at three error bounds, while also reporting the highest average end-to-end throughput among the evaluated compressors (676 GB/s compression, 785 GB/s decompression on GH200), with an additional A100 portability check. The authors argue that the mandatory two-pass variable-length-encoding architecture leaves arithmetic headroom in the rate pass, so richer prediction need not cost extra memory bandwidth.

Significance. If the results hold, FSZ is a valuable systems contribution: it demonstrates that adaptive multi-order prediction and cross-block state can be added to a pure-GPU, single-kernel error-bounded compressor without an end-to-end throughput regression. The evaluation is largely credible: external baselines (cuSZp-O/P, FZ-GPU, cuZFP), public datasets, a per-field ablation, a tile-configuration sweep, and a second-GPU portability check are all present. The primary weakness is that the paper's central mechanism — that the rate pass is bandwidth-bound and therefore the extra prediction arithmetic is free — is not directly measured; the reported end-to-end throughput could partly reflect shorter Phase-3 encode/decode work due to higher compression ratio rather than hidden arithmetic headroom. This does not invalidate the practical result, but it means the conceptual claim is stated more strongly than the evidence supports.

major comments (3)
  1. [§II-B, §V-C] The 'arithmetic headroom' claim is load-bearing for the title and abstract, but it is never measured directly. The paper asserts that the rate pass is memory-bandwidth-bound and therefore extra LZ2 state, tile-sum accumulation, and four-way cost evaluation hide inside memory stalls. However, the throughput comparison in Fig. 5 and Section V-C is end-to-end, and FSZ writes far fewer encoded bytes than cuSZp due to its higher CR, so its Phase 3 is shorter. The observed 'no throughput penalty' is therefore consistent with a rate pass that is actually compute-bound, with the slowdown offset by encode-side savings. The only direct evidence, the 5-read baseline in Fig. 6 being 2.8x slower, shows that extra global reads are expensive, not that extra ALU work is free. Please report per-phase timings (Phase 1 rate pass vs Phase 3 encode), register usage/occupancy, and small-input/latency-bound re
  2. [§IV-B4, §III-A] The claim 'no warp divergence' appears inconsistent with per-tile adaptive selection. Each thread in a warp processes a different tile and can choose a different LZ order/centering flag. If Phase 3 branches on the selected variant, threads in the same warp taking different paths will diverge. The text says 'all threads independently select their own tile’s variant with no divergent branches within a warp,' but independent selection does not by itself eliminate divergence. If the implementation uses predication, arithmetic selects, or computes all variants and copies the chosen one, the paper should state that explicitly. This matters because divergence-free adaptive selection is one of the listed throughput mechanisms in Observation II and Section III-A.
  3. [§V-C] The paper's headline throughput numbers are reported as single averages over 10 runs, with no standard deviation, min/max, or per-field spread. Given that the central claim is 'no throughput penalty' and that the paper presents averages such as 676 GB/s and 785 GB/s, the absence of any variance information makes it difficult for a reader to assess whether the differences from cuSZp-O/P are meaningful. Add at least standard deviations or error bars for the throughput results, and ideally report the number of runs in a way that allows the comparison to be evaluated.
minor comments (5)
  1. [§IV-B3, Eq. (4)] The statement that 'the LZ1 and LZ2 residuals with centering are identical to the no-centering residuals for all elements except the first 1–2 boundary elements per tile' should be made precise for i=0 and i=1. Equations (1)–(2) define LZ1 residuals only for i≥1; the LZ2 definition for i=0,1 is not given. Please define how the first one or two residuals are formed, since the centering correction for block 0 depends on that.
  2. [§V-D, Fig. 8] The observation that cuZFP exceeds the error bound at 23.0% of points is expected because cuZFP is a fixed-rate, non-error-bounded compressor. This is useful context, but the wording in Observation III could be read as a defect of cuZFP rather than an inherent property of the comparison. Rephrase to clarify that this is by design.
  3. [§III-A] The paper does not state whether source code or an artifact will be made available. For a systems paper making throughput and ratio claims, a code availability statement is helpful for reproducibility and would strengthen the contribution.
  4. [Table III] In Table III, single-field datasets (Truss) omit the standard deviation, but the table caption does not state this explicitly. Also, the footnote 'Best averages in bold blue' is acceptable, but the reader has to infer that bold means best across the four error-bounded compressors; consider stating that cuZFP is excluded because it is fixed-rate.
  5. [§VI-A] Figure 9 shows per-field CR for CESM-ATM at a single error bound; the text says this is representative. It would be useful to state explicitly that the per-field analysis at REL 1E-3 and 1E-4 shows the same qualitative pattern, or to include a supplementary figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FSZ's prediction-stage derivation is self-contained and externally benchmarked.

full rationale

The paper's core compression-ratio claims are established by a progressive, internally consistent design: the cross-block prediction state is defined by Eq. (2) as a carry of the previous block's last value, eliminating boundary residuals; the four-way evaluation rests on the elementary finite-difference identity in Eq. (4), which is proved in the text and does not presume the result it explains; and the ablation study (Table IV) separately attributes CR gains to each of the three mechanisms. The throughput claims are not derived from the arithmetic-headroom argument alone; they are measured as end-to-end benchmarks against external, other-group compressors (cuSZp-P, cuSZp-O, FZ-GPU, cuZFP) on public datasets and two GPU platforms. The central 'prediction-throughput trade-off broken' claim is therefore not forced by construction: the paper's own 5-read baseline demonstrates a bandwidth cost, and the decisive evidence is empirical comparison with baselines. The only tuned configuration (B=32, T=8) is disclosed with a full sweep in Table V that shows T=16 yields better CR but worse throughput, and the paper honestly selects the throughput-friendly setting. Self-citations in the reference list concern related collective-communication and wavelet work and are not load-bearing for FSZ's novel prediction mechanisms; no uniqueness theorem or prior-work-only justification is invoked to select FSZ's design. Thus no definitional, self-citational, renamed-result, or fit-called-prediction circularity is present.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

No new physical or mathematical entities are postulated. The tile mean μ, the four prediction variants, and the rate-header metadata are internal algorithmic structures of the compressor; they carry no independent-evidence requirement because they make no predictions outside the system. The only hand-chosen numbers are the tile configuration (swept and disclosed) and a minor write-skip threshold.

free parameters (2)
  • Tile configuration = B=32, T=8 (256-element tiles); sweep covered B=64/T=4, B=64/T=8, B=32/T=4, B=32/T=16
    Block size and blocks-per-tile are hand-chosen engineering knobs, validated by a 5-configuration sweep (Table V). The paper honestly reports that T=16 never loses on CR but keeps T=8 for throughput - disclosed tuning, not hidden fitting.
  • Pre-zero skip threshold = Compression ratio > 100
    Heuristic for skipping writes of zero-rate blocks (Section III-A). Minor implementation choice, not central to any claim.
assumptions (6)
  • standard math Constant offsets cancel exactly in k-th order finite differences for i >= k: Δ^k(q_i − μ) = Δ^k q_i
    Section IV-B3, Eq. (4). Elementary and correct; the enabler of single-pass four-way evaluation.
  • domain assumption The two-pass structure (rate pass then encode pass) is fundamental for single-kernel variable-length encoding
    Section II-A. Holds for this architecture; stated more strongly than necessary - single-pass variable-rate encoders with retry/feedback exist - but a fair description of the cuSZp design space.
  • domain assumption The rate pass is memory-bandwidth-bound with idle ALU capacity ("arithmetic headroom")
    Section II-B. Load-bearing for the no-throughput-penalty claim; supported only indirectly by measured throughput and the 5-read comparison, not by occupancy or register measurements.
  • domain assumption Quantization q_i = round(d_i/(2·eb)) guarantees pointwise error <= eb, and centering/prediction are exact invertible integer transforms
    Section IV-B2 (Error-bound preservation). Standard SZ-style quantization argument; correct given finite-precision handling.
  • domain assumption Rate formula r_b = ceil(log2(max|ℓ|+1)) with a 4-byte sign map models compressed block size exactly
    Section IV-B2, Eq. (3). Definition of the fixed-length encoding; consistent with the described bit-shuffled format.
  • domain assumption Benchmarks on one GH200 (CUDA 13.0) and one A100 generalize to the claimed performance classes
    Section V-A, V-D. Single-machine measurements; standard for the field but unverified by any third party.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FSZ: Breaking the Prediction-Throughput Trade-off in GPU Lossy Compression." pith.science (2026). https://pith.science/paper/63FOJIPL

@misc{pith2026260715413,
  author       = {Pith},
  title        = {Pith review of: FSZ: Breaking the Prediction-Throughput Trade-off in GPU Lossy Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63FOJIPL}},
  note         = {Machine review of arXiv:2607.15413}
}
read the original abstract

Existing fast GPU error-bounded lossy compressors have achieved high throughput through pure-GPU single-kernel designs, but their compression ratios remain limited because they typically apply a fixed first-order predictor on independent blocks. We propose FSZ, a GPU error-bounded lossy compressor that redesigns the prediction stage with three mutually reinforcing algorithmic innovations to achieve both higher compression ratios and higher throughput within a single CUDA kernel: (1) cross-block prediction state carries Lorenzo prediction state across block boundaries within 256-element tiles, eliminating 7 out of 8 boundary residuals that inflate encoding rates; (2) per-tile adaptive multi-order prediction and centering adaptively selects the best compression strategy per tile from first-order, second-order, and centering variants; and (3) a single-pass four-way evaluation exploits a mathematical property of finite differences to evaluate all variants from a single data read, enabling richer prediction within the same bandwidth budget as a fixed predictor. Experiments on NVIDIA GH200 GPU with 8 real-world application datasets show that 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. Notably, these gains come with no throughput penalty: FSZ simultaneously achieves the highest average throughput (676 GB/s compression, 785 GB/s decompression) among all evaluated compressors.

Figures

Figures reproduced from arXiv: 2607.15413 by the authors.

Figure 1
Figure 1. Overview of the FSZ compression (top) and decompression [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Cross-block prediction state. (a) Independent blocks: each [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Adaptive multi-order prediction on Block 1 with cross-block [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Compression (top) and decompression (bottom, mirrored) throughput at three error bounds per dataset. FSZ (solid blue) consistently [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Single-pass four-way evaluation (blue) vs. naive 5-read [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Rate-distortion curves (PSNR vs bit rate) for FSZ and baseline compressors on 6 representative datasets. Higher and further left is [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visual quality on CESM-ATM RELHUM at matched CR (FSZ 14.88, cuZFP 14.86). Each panel shows the full slice with a [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Per-field CR comparison on all 33 CESM-ATM fields at REL 1E-2, sorted by FSZ’s advantage over cuSZp-O. FSZ achieves up to [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 11
Figure 11. Figure 11: Compression and decompression throughput on NVIDIA [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 10
Figure 10. Figure 10: Throughput of 5 tile configurations at 3 error bounds. Default [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 5 canonical work pages

  1. [1]

    cuszp2: A gpu lossy compressor with extreme throughput and optimized compression ratio,

    Y . Huang, S. Di, G. Li, and F. Cappello, “cuszp2: A gpu lossy compressor with extreme throughput and optimized compression ratio,” inProceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis, ser. SC ’24, 2024, DOI: https://doi.org/10.1109/SC41406.2024.00021

    arXiv 2024

  2. [2]

    SDRBench: Scientific data reduction benchmark for lossy compressors,

    K. Zhao, S. Di, X. Lian, S. Li, D. Tao, J. Bessac, Z. Chen, and F. Cappello, “SDRBench: Scientific data reduction benchmark for lossy compressors,” in2020 IEEE International Conference on Big Data (Big Data), 2020, pp. 2716–2724, DOI: https://doi.org/10.1109/ BigData50022.2020.9378449

    arXiv 2020

  3. [3]

    Zccl: Significantly improving col- lective communication with error-bounded lossy compression,

    J. Huang, S. Di, X. Yu, Y . Zhai, Z. Zhang, J. Liu, X. Lu, K. Raffenetti, H. Zhou, K. Zhao, K. Alharthi, Z. Chen, F. Cap- pello, Y . Guo, and R. Thakur, “Zccl: Significantly improving col- lective communication with error-bounded lossy compression,”arXiv preprint arXiv:2502.18554, 2025, DOI: https://doi.org/10.48550/arXiv. 2502.18554

    arXiv doi:10.48550/arxiv.2502.18554 2025

  4. [4]

    An optimized error-controlled mpi collective framework integrated with lossy compression,

    J. Huang, S. Di, X. Yu, Y . Zhai, Z. Zhang, J. Liu, X. Lu, K. Raffenetti, H. Zhou, K. Zhao, Z. Chen, F. Cappello, Y . Guo, and R. Thakur, “An optimized error-controlled mpi collective framework integrated with lossy compression,” in2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2024, pp. 752–764, DOI: https://doi.org/10.110...

    arXiv 2024

  5. [5]

    gzccl: Compression-accelerated collective communication framework for gpu clusters,

    J. Huang, S. Di, X. Yu, Y . Zhai, J. Liu, Y . Huang, K. Raffenetti, H. Zhou, K. Zhao, X. Lu, Z. Chen, F. Cappello, Y . Guo, and R. Thakur, “gzccl: Compression-accelerated collective communication framework for gpu clusters,” inProceedings of the 38th ACM International Conference on Supercomputing, ser. ICS ’24, 2024, p. 437–448, DOI: https://doi.org/ 10.1...

    arXiv 2024

  6. [6]

    ghzccl: Advancing gpu-aware collective communications with homo- morphic compression,

    J. Huang, S. Di, Y . Huang, Z. Chen, F. Cappello, Y . Guo, and R. Thakur, “ghzccl: Advancing gpu-aware collective communications with homo- morphic compression,” inProceedings of the 39th ACM International Conference on Supercomputing, ser. ICS ’25, 2025, pp. 43–56, DOI: https://doi.org/10.1145/3721145.3733642

    arXiv 2025

  7. [7]

    Fast error-bounded lossy hpc data compression with sz,

    S. Di and F. Cappello, “Fast error-bounded lossy hpc data compression with sz,” in2016 IEEE International Parallel and Distributed Processing Symposium (IPDPS). IEEE, 2016, pp. 730–739, DOI: https://doi.org/ 10.1109/IPDPS.2016.11

    doi:10.1109/ipdps.2016.11 2016

  8. [8]

    Significantly improving lossy compression for scientific data sets based on multidimensional prediction and error-controlled quantization,

    D. Tao, S. Di, Z. Chen, and F. Cappello, “Significantly improving lossy compression for scientific data sets based on multidimensional prediction and error-controlled quantization,” in2017 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2017, pp. 1129–1139, DOI: https://doi.org/10.1109/IPDPS.2017.115

    work page doi:10.1109/ipdps.2017.115 2017

Show all 36 references
  1. [9]

    Fixed-rate compressed floating-point arrays,

    P. Lindstrom, “Fixed-rate compressed floating-point arrays,”IEEE Transactions on Visualization and Computer Graphics, vol. 20, pp. 2674–2683, 2014, DOI: https://doi.org/10.1109/TVCG.2014.2346458

    2014

  2. [10]

    cuszp: An ultra-fast gpu error-bounded lossy compression framework with optimized end-to- end performance,

    Y . Huang, S. Di, X. Yu, G. Li, and F. Cappello, “cuszp: An ultra-fast gpu error-bounded lossy compression framework with optimized end-to- end performance,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ser. SC...

    2023

  3. [11]

    Fz-gpu: A fast and high-ratio lossy compressor for scientific com- puting applications on gpus,

    B. Zhang, J. Tian, S. Di, X. Yu, Y . Feng, X. Liang, D. Tao, and F. Cap- pello, “Fz-gpu: A fast and high-ratio lossy compressor for scientific com- puting applications on gpus,” inProceedings of the 32nd International Symposium on High-Performance Parallel and Distributed Comp...

    2023

  4. [12]

    Single-pass parallel prefix scan with decoupled look-back,

    M. G. Duane Merrill, “Single-pass parallel prefix scan with decoupled look-back,” https://research.nvidia.com/sites/default/files/pubs/2016-03 Single-pass-Parallel-Prefix/nvr-2016-002.pdf, 2016

    2016

  5. [13]

    GPU lossy compression for HPC can be versatile and ultra-fast,

    Y . Huang, S. Di, G. Li, and F. Cappello, “GPU lossy compression for HPC can be versatile and ultra-fast,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ser. SC ’25, 2025, DOI: https://doi.org/10.1145/3712285. 3759817

    2025 doi

  6. [14]

    Open scientific visualization datasets,

    Pavol Klacansky, “Open scientific visualization datasets,” http:// klacansky.com/open-scivis-datasets/, online

  7. [15]

    The Community Earth System Model (CESM) large ensemble project: A community resource for studying climate change in the presence of internal climate variability,

    J. E. Kayet al., “The Community Earth System Model (CESM) large ensemble project: A community resource for studying climate change in the presence of internal climate variability,”Bulletin of the American Meteorological Society, vol. 96, no. 8, pp. 1333–1349, 2015, DOI: https:...

    2015 doi

  8. [16]

    EXAALT: Exascale atomistic capability for accuracy, length, and time,

    The EXAALT team, “EXAALT: Exascale atomistic capability for accuracy, length, and time,” https://www.exascaleproject.org/ research-project/exaalt/, 2018, online

    2018

  9. [17]

    Hacc: Simulating sky surveys on state-of-the-art supercomputing architectures,

    S. Habib, A. Pope, H. Finkel, N. Frontiere, K. Heitmann, D. Daniel, P. Fasel, V . Morozov, G. Zagaris, T. Peterkaet al., “Hacc: Simulating sky surveys on state-of-the-art supercomputing architectures,”New As- tronomy, vol. 42, pp. 49–65, 2016, DOI: https://doi.org/10.1016/j.ne...

    2016 doi

  10. [18]

    Hurricane ISABEL simulation data, http://vis.computer.org/ vis2004contest/data.html, 2004, online

    2004

  11. [19]

    The mixing transition in Rayleigh-Taylor instability,

    A. W. Cook, W. Cabot, and P. L. Miller, “The mixing transition in Rayleigh-Taylor instability,”Journal of Fluid Mechanics, vol. 511, pp. 333–362, 2004, DOI: https://doi.org/10.1017/S0022112004009681

    2004 doi

  12. [20]

    NYX simulation, https://amrex-astro.github.io/Nyx, 2019, online

    2019

  13. [21]

    Virtual inspection of ad- ditively manufactured parts,

    P. Klacansky, H. Miao, A. Gyulassy, A. Townsend, K. Champley, J. Tringe, V . Pascucci, and P.-T. Bremer, “Virtual inspection of ad- ditively manufactured parts,” in2022 IEEE 15th Pacific Visualization Symposium (PacificVis), 2022, pp. 81–90, DOI: https://doi.org/10.1109/ Pacif...

    2022

  14. [22]

    The near-real-time SCALE- LETKF system: A case of the september 2015 Kanto-Tohoku heavy rainfall,

    G.-Y . Lien, T. Miyoshi, S. Nishizawa, R. Yoshida, H. Yashiro, S. A. Adachi, T. Yamaura, and H. Tomita, “The near-real-time SCALE- LETKF system: A case of the september 2015 Kanto-Tohoku heavy rainfall,”SOLA, vol. 13, pp. 1–6, 2017, DOI: https://doi.org/10.2151/ sola.2017-001

    2015

  15. [23]

    Lindstrom, “cuZFP,” https://github.com/LLNL/zfp/tree/develop/src/ cuda zfp

    P. Lindstrom, “cuZFP,” https://github.com/LLNL/zfp/tree/develop/src/ cuda zfp

  16. [24]

    Error-controlled lossy compression optimized for high compression ratios of scientific datasets,

    X. Liang, S. Di, D. Tao, S. Li, S. Li, H. Guo, Z. Chen, and F. Cappello, “Error-controlled lossy compression optimized for high compression ratios of scientific datasets,” in2018 IEEE International Conference on Big Data (Big Data), 2018, pp. 438–447, DOI: https://doi.org/10.1...

    2018

  17. [25]

    Sig- nificantly improving lossy compression for hpc datasets with second- order prediction and parameter optimization,

    K. Zhao, S. Di, X. Liang, S. Li, D. Tao, Z. Chen, and F. Cappello, “Sig- nificantly improving lossy compression for hpc datasets with second- order prediction and parameter optimization,” inProceedings of the 29th International Symposium on High-Performance Parallel and Distri...

    2020 doi

  18. [27]

    Exploring wavelet transform usages for error-bounded scientific data compression,

    J. Huang, J. Liu, S. Di, Y . Zhai, Z. Jian, S. Wu, K. Zhao, Z. Chen, Y . Guo, and F. Cappello, “Exploring wavelet transform usages for error-bounded scientific data compression,” in2023 IEEE International Conference on Big Data (BigData), 2023, pp. 4233–4239, DOI: https: //doi...

    2023

  19. [28]

    Lossy scientific data compression with SPERR,

    S. Li, P. Lindstrom, and J. Clyne, “Lossy scientific data compression with SPERR,” in2023 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2023, pp. 1007–1017, DOI: https: //doi.org/10.1109/IPDPS54959.2023.00104

    2023

  20. [29]

    Towards improving reverse time migration performance by high-speed lossy compression,

    Y . Huang, K. Zhao, S. Di, G. Li, M. Dmitriev, T.-L. D. Tonellot, and F. Cappello, “Towards improving reverse time migration performance by high-speed lossy compression,” in2023 IEEE/ACM 23rd International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2023, pp. ...

    2023

  21. [30]

    Full-state quantum circuit simulation by using data compression,

    X.-C. Wu, S. Di, E. M. Dasgupta, F. Cappello, H. Finkel, Y . Alexeev, and F. T. Chong, “Full-state quantum circuit simulation by using data compression,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ser. SC ’19...

    2019

  22. [31]

    Exploration of lossy compression for application-level checkpoint/restart,

    N. Sasaki, K. Sato, T. Endo, and S. Matsuoka, “Exploration of lossy compression for application-level checkpoint/restart,” in2015 IEEE International Parallel and Distributed Processing Symposium, 2015, pp. 914–922, DOI: https://doi.org/10.1109/IPDPS.2015.67

    2015 doi

  23. [32]

    LLaMA: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample, “LLaMA: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023, DOI: https://...

    2023 doi arXiv

  24. [33]

    High-speed data communication with advanced networks in large language model training,

    L. Dai, H. Qi, W. Chen, and X. Lu, “High-speed data communication with advanced networks in large language model training,”IEEE Micro, vol. 44, no. 2, p. 31–40, Jan. 2024. [Online]. Available: https://doi.org/10.1109/MM.2024.3360081

    2024

  25. [34]

    cuSZ: An efficient gpu-based error-bounded lossy compression framework for scientific data,

    J. Tian, S. Di, K. Zhao, C. Rivera, M. H. Fulp, R. Underwood, S. Jin, X. Liang, J. Calhoun, D. Tao, and F. Cappello, “cuSZ: An efficient gpu-based error-bounded lossy compression framework for scientific data,” inProceedings of the ACM International Conference on Parallel Arch...

    2020

  26. [35]

    Ultrafast error-bounded lossy compression for scientific datasets,

    X. Yu, S. Di, K. Zhao, J. Tian, D. Tao, X. Liang, and F. Cappello, “Ultrafast error-bounded lossy compression for scientific datasets,” in Proceedings of the 31st International Symposium on High-Performance Parallel and Distributed Computing, 2022, DOI: https://doi.org/10.1145...

    2022

  27. [36]

    Boosting scientific error-bounded lossy compression through optimized synergistic lossy-lossless orchestration,

    S. Wu, J. Pan, J. Liu, J. Tian, Z. Qiu, J. Huang, K. Zhao, X. Liang, S. Di, Z. Chen, and F. Cappello, “Boosting scientific error-bounded lossy compression through optimized synergistic lossy-lossless orchestration,” inProceedings of the International Conference for High Perfor...

    2025

  28. [37]

    Fast and effective lossy compression on GPUs and CPUs with guaranteed error bounds,

    A. Fallin, N. Azami, S. Di, F. Cappello, and M. Burtscher, “Fast and effective lossy compression on GPUs and CPUs with guaranteed error bounds,” in2025 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2025, pp. 874–887, DOI: https://doi.org/10.1109/ IP...

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.