{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.17379", "kind": "paper", "url": "https://arxiv.org/abs/2608.17379"}, {"id": "gh:zhang677/ptxbench", "kind": "repo", "url": "https://github.com/zhang677/PTXBench"}, {"id": "hf:accrl/accrl-training", "kind": "dataset", "url": "https://huggingface.co/datasets/AccRL/accrl-training"}], "authors": ["Genghan Zhang", "Yixin Dong", "Chengze Fan", "Zhichen Zeng", "Yueming Yuan", "Shaowei Zhu", "Kunle Olukotun"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.17379", "source_url": "https://arxiv.org/abs/2608.17379"}, "categories": ["Architecture-specific programming", "Code", "Code & Software", "Code generation", "GPU kernel optimization", "Iterative code repair", "Software & AI Compute"], "description": {"en": "PTXBench evaluates LLMs in generating architecture-specific PTX for GPU kernel optimization. It measures functional correctness, execution of target instructions, and speedup over frontier libraries across GEMM and attention workloads on H100 and B200 GPUs."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.17379", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.17379", "title": "PTXBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Stanford University", "Carnegie Mellon University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.17379", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.17379", "source_url": "https://openalex.org/W7203748250"}, "key": "claire-radar:2608.17379", "modality": null, "name": "PTXBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_ptxbench_11575102", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "fb2f06189d3d8055cc09f950851b31fc9905ebff2f8d291a959fd8ff2d5cebcc", "relation": "introduces", "review_state": "source-reviewed", "reviewed_at": "2026-08-20", "source_url": "https://arxiv.org/abs/2608.17379"}, "publisher": null, "released": "2026-08-18", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.17379", "source_url": "https://arxiv.org/abs/2608.17379"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-17379", "source": "claire_radar", "source_benchmark_id": "2608.17379", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Software & AI Compute"], "area": "Code & Software", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/zhang677/PTXBench", "githubScope": "benchmark_repo", "githubStars": 31, "hfDailySubmittedAt": null, "hfDataset": "AccRL/accrl-training", "hfDatasetDownloads": 288, "hfDatasetLikes": 0, "hfPaperUpvotes": 4, "hfPaperUrl": "https://huggingface.co/papers/2608.17379", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-09-30", "state": "stale"}, "hfDatasetLikes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-09-30", "state": "stale"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-09-30", "state": "stale"}}, "status": "fresh"}, "availability": {"evaluatorStatus": "available", "githubStatus": "available", "hfDatasetStatus": "available", "paperStatus": "available", "submissionStatus": "not_found"}, "capabilities": ["Code generation", "GPU kernel optimization", "Architecture-specific programming", "Iterative code repair"], "confidence": "High", "construction": "Unknown", "constructionDetail": "PTXBench tests architecture-specific GPU kernel generation on H100 and B200 hardware, including correctness, required-instruction execution and speedup.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "973bda730a80d543c43a5fee47f05df482a9a432407dfbc4a14ec612f974a73a", "model": "deepseek-v4-flash"}, "curation": {"reviewedAt": "2026-08-20", "sources": ["https://arxiv.org/abs/2608.17379", "https://arxiv.org/html/2608.17379", "https://github.com/zhang677/PTXBench", "https://huggingface.co/datasets/AccRL/accrl-training"], "state": "source-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "PTXBench evaluates LLMs in generating architecture-specific PTX for GPU kernel optimization. It measures functional correctness, execution of target instructions, and speedup over frontier libraries across GEMM and attention workloads on H100 and B200 GPUs.", "detail": {"protocol": {"language": "CUDA C++ with inline PTX", "primaryMetric": "Correctness, target-instruction correctness, best speedup and Fast_p", "tasks": "5 primary BF16 workloads plus generalization workloads", "version": "arXiv v1"}, "taskBreakdown": ["GEMM", "Multi-head attention forward", "Causal attention forward", "Multi-head attention backward", "Causal attention backward"]}, "domainCuration": {"method": "primary-source-review-v1", "state": "reviewed"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "We introduce PTXBench, a benchmark for evaluating and adapting large language models (LLMs) to use architecture-specific PTX for GPU kernel optimization."}, "familyId": "bmf_dc9322ec40cc", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/zhang677/PTXBench", "url": "https://github.com/zhang677/PTXBench"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/zhang677/PTXBench", "url": "https://huggingface.co/papers/2608.17379"}, {"basis": "record-resource-link", "linkedUrl": "https://github.com/zhang677/PTXBench", "url": "https://github.com/zhang677/PTXBench"}], "scope": "benchmark_repo", "url": "https://github.com/zhang677/PTXBench", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_ptxbench_11575102", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Software & Cloud"], "links": {"code": "https://github.com/zhang677/PTXBench", "data": "https://huggingface.co/datasets/AccRL/accrl-training", "hfPaper": "https://huggingface.co/papers/2608.17379", "paper": "https://arxiv.org/abs/2608.17379", "pdf": "https://arxiv.org/pdf/2608.17379", "project": "https://github.com/zhang677/PTXBench", "report": "https://arxiv.org/abs/2608.17379"}, "metrics": [], "motivation": "We introduce PTXBench, a benchmark for evaluating and adapting large language models (LLMs) to use architecture-specific PTX for GPU kernel optimization.", "name": "PTXBench", "oneLine": "Evaluates whether LLM agents can generate correct, architecture-specific GPU kernels that execute required PTX instructions and outperform libraries.", "paperTitle": "PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX", "primaryDomain": "Software & AI Compute", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.17379", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "PTXBench Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/zhang677/PTXBench"}], "ranking": {"90d": {"confidence": "High", "coverage": 1.0, "datasetDownloadRank": 38, "datasetRankPopulation": 85, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.46261344253546793, "value": 31}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": 0.5835281833665731, "value": 288}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.2843790034276272, "value": 4}}, "confidence": "High", "coverage": 1.0, "rank": 89, "score": 47}, "method": "attention-ranking-v13", "rank": 89, "score": 47}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-18", "source": {"authors": ["Genghan Zhang", "Yixin Dong", "Chengze Fan", "Zhichen Zeng", "Yueming Yuan", "Shaowei Zhu", "Kunle Olukotun"], "categories": ["cs.CL", "cs.AI"], "id": "2608.17379", "title": "PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX", "type": "arxiv", "url": "https://arxiv.org/abs/2608.17379"}, "sourceUpdatedAt": "2026-08-18T05:14:41Z", "topics": ["Code"], "venueAttempts": [], "whyItMatters": "Addresses the lack of standardized evaluation for LLM-driven GPU kernel optimization with architecture-specific PTX, providing a reproducible testbed to compare model capabilities and guide improvements in exploiting evolving GPU architectures."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
