Researchers have evaluated seven different strategies for generating test inputs to find bugs in tensor kernels, which are crucial for AI and machine learning computations. Using a seeded fuzzer on an RTX 3060 GPU, they found that boundary-only shape sampling was the most effective and safe strategy, achieving 78% bug recall with zero false positives on correct kernels. While adversarial value sampling could achieve higher recall, it significantly increased false positives due to the injection of NaN and Inf values. AI
IMPACT Improves the reliability of core AI/ML computation kernels, potentially reducing bugs in deployed models.
RANK_REASON Academic paper detailing a new method for testing software components. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
