Detection Report: Out-of-Memory Graph Processing Acceleration via Algorithmic-Hardware Codesign on FPGAs

Academic fraud report · Geng Detector

Summary

This report evaluates a 2026 IEEE Transactions on Computers paper (Author's Accepted Version) on FPGA-based out-of-memory graph processing. The overall verdict is highly suspicious. The most serious issue is a critical data inconsistency: the abstract claims only a 1.3x device-level energy-efficiency improvement versus a GPU baseline, while the conclusion suddenly asserts a 20.8x overall energy-efficiency improvement over the ReGraph-OoM FPGA baseline. The body (Section VI-E) provides no derivation or calculation supporting the 20.8x figure, suggesting copy-paste from another draft or paper. Several secondary concerns indicate insufficient proofreading: the LaTeX header still reads 'JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021'; an acknowledgments sentence contains a truncated grant number ('Grant and 62172380'); and Algorithm 1, Line 5, calls GETBOTTLENECKTYPE without arguments. These are individually minor but collectively suggest hasty, fragmented manuscript assembly. Numeric consistency within tables (e.g., 8236.5/34W ≈ 242.25 MTEPS/W) appears internally coherent, with no evidence of fabricated tabular data. Limitations: image-level reuse and splicing checks were not performed; conclusions are tentative and require original figures and experimental logs for confirmation.

Verdict

🟠 Highly suspicious. The verdict rests primarily on a major internal contradiction between the abstract and conclusion regarding energy-efficiency gains, supported by multiple proofreading failures. A definitive determination of misconduct requires the authors' original experimental logs and the publisher's investigation.

Key findings

Core data contradiction (Finding 1): The abstract reports only a 1.3x device-level energy-efficiency improvement versus a GPU solution, while the conclusion claims a 20.8x overall energy-efficiency improvement over the ReGraph-OoM FPGA baseline. No derivation or supporting calculation appears in Section VI-E, indicating possible copy-paste error across drafts.
Unmodified LaTeX template (Finding 2): Every page header reads 'JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021', inconsistent with a 2026 submission to IEEE Transactions on Computers.
Truncated grant number (Finding 3): The acknowledgments state '...under Grant and 62172380...', with an apparent missing grant number between 'Grant' and 'and'.
Incomplete pseudocode call (Finding 4): Algorithm 1, Line 5, bottleneck ← GETBOTTLENECKTYPE omits required arguments present in the preceding line (e.g., Ccurt, H).
Image analysis not performed (Finding 5): Pixel-level checks for figure reuse or splicing were not feasible from text extraction. Tabular arithmetic (e.g., Table IV: 8236.5 / 34W ≈ 242.25 MTEPS/W) is internally consistent.

Evidence highlights

DOI: 10.1109/TC.2026.3701992
Abstract energy-efficiency claim: 1.3x vs. GPU (device-level energy efficiency).
Conclusion energy-efficiency claim: 20.8x overall vs. ReGraph-OoM (FPGA baseline); 1.3x vs. GPU.
Missing derivation location: Section VI-E contains no derivation, table, or figure supporting the 20.8x value.
Header text: 'JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021' on every page.
Acknowledgments fragment: 'National Natural Science Foundation of China under Grant and 62172380'.
Algorithm 1, Line 5: bottleneck ← GETBOTTLENECKTYPE (no arguments passed).
Table IV arithmetic check: 8236.5 MTEPS / 34 W ≈ 242.25 MTEPS/W — internally consistent.

Notes

The combination of a glaring numerical inconsistency and repeated low-level proofreading failures raises concern about manuscript assembly practices rather than deliberate fabrication alone.
Recommended follow-up actions: request authors' raw experimental logs to substantiate the 20.8x figure; flag the abstract–conclusion discrepancy on PubPeer; require corrections to template, grant number, and pseudocode before final publication.
This report is AI-assisted and intended for academic discussion; final determinations of misconduct rest with the publisher and relevant institutional bodies. False positives and false negatives are possible.