chipintegrity.org GitHub
ABOUT

Research, standards and how to cite

The checks build on published work in numerical analysis and hardware reliability, and write their results in an open industry format. This page lists that work, who maintains the project, and how to cite it.

ON THIS PAGE YOU WILL LEARN
  1. The papers and standards the checks rely on
  2. Who maintains the project
  3. How to cite the tool
TAKE PART

Research and standards this builds on

  • Zheng et al., OSDI 2026. SDCs in the Wild: characterising and diagnosing SDC-defective GPUs in production LLM training. The source of the finding that microbenchmark screening misses most faulty units. usenix.org
  • Lei et al., OSDI 2026. Safeguarding LLM Training at Scale: online SDC detection and insights from 35 million GPU hours. usenix.org
  • Tung, Huang, Saxena, Shirvani, Hukerikar, Jain, Tyagi and Gongalore, DSN 2026. The Anatomy of Silent Data Corruption: GPU error pattern study and modeling guidance. Gate-level fault injection on a data-center GPU: half of all corruptions are zeroed values, under 40% of bit-flip corruptions are single-bit, and corruptions repeat at warp-aligned spacings. The self-test's gpu-sdc faults follow its measurements. arxiv.org
  • Higham, Accuracy and Stability of Numerical Algorithms. The componentwise rounding error bound for inner products used by the reference check.
  • Huang and Abraham, 1984. Algorithm-based fault tolerance for matrix operations, IEEE Transactions on Computers. The row and column checksums behind probe 6.
  • Fasi, Higham, Mikaitis and Pranesh, 2021. Numerical behavior of NVIDIA tensor cores. The measurement that tensor cores accumulate with truncation, which sets the fp16 and bf16 bound. peerj.com
  • DeepSeek-AI, 2024. DeepSeek-V3 technical report. Reports that Hopper's FP8 tensor cores keep 14 bits when they add, which is why the FP8 inputs here are chosen so every sum fits in 13. arxiv.org
  • Valpey and Pai, 2025. Modeling three generations of tensor cores. A formal model of rounding and accumulation order on Volta, Turing and Ampere. arxiv.org
  • OCP Test and Validation. The open diagnostic output format and reference implementation all six probes write to. ocp-diag-core
TAKE PART

About and cite

Chip Integrity is maintained by Yasha Khandelwal, Bengaluru. The code is MIT licensed. The project stays vendor neutral: the same tests run on every chip, and a row is published only with its result file.

Yasha Khandelwal. ai-chip-integrity: an open tester for silent computation errors in AI chips.
Version 0.4.2, 2026. https://doi.org/10.5281/zenodo.23393790
https://github.com/tech4biz-yasha/ai-chip-integrity

Questions, rows and findings: yasha.khandelwal@tech4biz.io or an issue on GitHub.