{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ATFTBZLXXY5EQJO4QD5DIIW2KQ","short_pith_number":"pith:ATFTBZLX","schema_version":"1.0","canonical_sha256":"04cb30e577be3a4825dc80fa3422da5405e42d1c72f0dd911608c521718e6fe2","source":{"kind":"arxiv","id":"2512.02551","version":3},"attestation_state":"computed","paper":{"title":"CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Albert Wang, Chris Shum, Guoyin Wang, Jiwei Li, Songqiao Su, Xiaoya Li","submitted_at":"2025-12-02T09:20:15Z","abstract_excerpt":"In this paper, we propose CUDA-L2, a system that combines large language models (LLMs) and reinforcement learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. Using CUDA execution speed as the RL reward, CUDA-L2 automatically optimizes HGEMM kernels across 1,000 configurations. CUDA-L2 systematically outperforms major matmul baselines to date, from the widely-used torch.matmul to state-of-the-art Nvidia's closed-source libraries, i.e., cuBLAS, cuBLASLt. In offline mode, where kernels are executed consecutively without time intervals, CUDA-L2 yield"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2512.02551","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-12-02T09:20:15Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"ad00fa6fe9fe6fcf025096a2c66c8ec8381423ee7d918eb8331cc0bfe86635d0","abstract_canon_sha256":"094b84eb51a6aec703ac377fcb82a8722c52654abba1eecaa92e8e83d95454e0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-14T01:20:49.283837Z","signature_b64":"FcK7Bcy6yFNfYWJJB32dLRrqJ6OW8fcG9IZuIyD575aC9EuOawHxkUCLavzslmsYD95zhkQpqcmAIN0N4h61Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"04cb30e577be3a4825dc80fa3422da5405e42d1c72f0dd911608c521718e6fe2","last_reissued_at":"2026-07-14T01:20:49.282920Z","signature_status":"signed_v1","first_computed_at":"2026-07-14T01:20:49.282920Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Albert Wang, Chris Shum, Guoyin Wang, Jiwei Li, Songqiao Su, Xiaoya Li","submitted_at":"2025-12-02T09:20:15Z","abstract_excerpt":"In this paper, we propose CUDA-L2, a system that combines large language models (LLMs) and reinforcement learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. Using CUDA execution speed as the RL reward, CUDA-L2 automatically optimizes HGEMM kernels across 1,000 configurations. CUDA-L2 systematically outperforms major matmul baselines to date, from the widely-used torch.matmul to state-of-the-art Nvidia's closed-source libraries, i.e., cuBLAS, cuBLASLt. In offline mode, where kernels are executed consecutively without time intervals, CUDA-L2 yield"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2512.02551","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2512.02551/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2512.02551","created_at":"2026-07-14T01:20:49.283321+00:00"},{"alias_kind":"arxiv_version","alias_value":"2512.02551v3","created_at":"2026-07-14T01:20:49.283321+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.02551","created_at":"2026-07-14T01:20:49.283321+00:00"},{"alias_kind":"pith_short_12","alias_value":"ATFTBZLXXY5E","created_at":"2026-07-14T01:20:49.283321+00:00"},{"alias_kind":"pith_short_16","alias_value":"ATFTBZLXXY5EQJO4","created_at":"2026-07-14T01:20:49.283321+00:00"},{"alias_kind":"pith_short_8","alias_value":"ATFTBZLX","created_at":"2026-07-14T01:20:49.283321+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":6,"sample":[{"citing_arxiv_id":"2605.25954","citing_title":"Step-TP: A Grounded, Step-Level Dataset with Chain-of-Thought Reasoning for LLM-Guided Tensor Program Optimization","ref_index":36,"is_internal_anchor":true},{"citing_arxiv_id":"2603.23566","citing_title":"AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19269","citing_title":"CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17978","citing_title":"AutoVecCoder: Teaching LLMs to Generate Explicitly Vectorized Code","ref_index":89,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19269","citing_title":"CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2603.28342","citing_title":"Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization","ref_index":23,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ","json":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ.json","graph_json":"https://pith.science/api/pith-number/ATFTBZLXXY5EQJO4QD5DIIW2KQ/graph.json","events_json":"https://pith.science/api/pith-number/ATFTBZLXXY5EQJO4QD5DIIW2KQ/events.json","paper":"https://pith.science/paper/ATFTBZLX"},"agent_actions":{"view_html":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ","download_json":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ.json","view_paper":"https://pith.science/paper/ATFTBZLX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2512.02551&json=true","fetch_graph":"https://pith.science/api/pith-number/ATFTBZLXXY5EQJO4QD5DIIW2KQ/graph.json","fetch_events":"https://pith.science/api/pith-number/ATFTBZLXXY5EQJO4QD5DIIW2KQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ/action/storage_attestation","attest_author":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ/action/author_attestation","sign_citation":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ/action/citation_signature","submit_replication":"https://pith.science/pith/ATFTBZLXXY5EQJO4QD5DIIW2KQ/action/replication_record"}},"created_at":"2026-07-14T01:20:49.283321+00:00","updated_at":"2026-07-14T01:20:49.283321+00:00"}