{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:UQKLUT5OACZ5Y4RG43GMMQZULG","short_pith_number":"pith:UQKLUT5O","schema_version":"1.0","canonical_sha256":"a414ba4fae00b3dc7226e6ccc64334598e87d37e5ff1092aa3c114c8531c7c68","source":{"kind":"arxiv","id":"2407.08608","version":2},"attestation_state":"computed","paper":{"title":"FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Ganesh Bikshandi, Jay Shah, Pradeep Ramani, Tri Dao, Vijay Thakkar, Ying Zhang","submitted_at":"2024-07-11T15:44:48Z","abstract_excerpt":"Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. FlashAttention elaborated an approach to speed up attention on GPUs through minimizing memory reads/writes. However, it has yet to take advantage of new capabilities present in recent hardware, with FlashAttention-2 achieving only 35% utilization on the H100 GPU. We develop three main techniques to speed up attention on Hopper GPUs: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-special"},"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":"2407.08608","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-07-11T15:44:48Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"7ccaf2708ba092d27a2f6b45130691fbbd7fbe07983a8bf8e1755befba2fd970","abstract_canon_sha256":"05f2351140b0c95c75fb9dd0b5482630db4c8fc20fffd2e571fba11022538a00"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T19:38:40.312261Z","signature_b64":"4kFCTH+VbKtdU1jZ/MKcyd8zWXktIkWOTKUSi+8eofNDvKZaC/wf288gJPoY96kzFBDSM/7i3FKJbi25DK3EAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a414ba4fae00b3dc7226e6ccc64334598e87d37e5ff1092aa3c114c8531c7c68","last_reissued_at":"2026-05-20T19:38:40.310580Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T19:38:40.310580Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Ganesh Bikshandi, Jay Shah, Pradeep Ramani, Tri Dao, Vijay Thakkar, Ying Zhang","submitted_at":"2024-07-11T15:44:48Z","abstract_excerpt":"Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. FlashAttention elaborated an approach to speed up attention on GPUs through minimizing memory reads/writes. However, it has yet to take advantage of new capabilities present in recent hardware, with FlashAttention-2 achieving only 35% utilization on the H100 GPU. We develop three main techniques to speed up attention on Hopper GPUs: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-special"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.08608","kind":"arxiv","version":2},"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/2407.08608/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":"2407.08608","created_at":"2026-05-20T19:38:40.310648+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.08608v2","created_at":"2026-05-20T19:38:40.310648+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.08608","created_at":"2026-05-20T19:38:40.310648+00:00"},{"alias_kind":"pith_short_12","alias_value":"UQKLUT5OACZ5","created_at":"2026-05-20T19:38:40.310648+00:00"},{"alias_kind":"pith_short_16","alias_value":"UQKLUT5OACZ5Y4RG","created_at":"2026-05-20T19:38:40.310648+00:00"},{"alias_kind":"pith_short_8","alias_value":"UQKLUT5O","created_at":"2026-05-20T19:38:40.310648+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":24,"internal_anchor_count":24,"sample":[{"citing_arxiv_id":"2409.01143","citing_title":"HexiScale: Facilitating Large Language Model Training over Heterogeneous Hardware","ref_index":49,"is_internal_anchor":true},{"citing_arxiv_id":"2511.02043","citing_title":"Flashlight: PyTorch Compiler Extensions to Accelerate Attention Variants","ref_index":16,"is_internal_anchor":true},{"citing_arxiv_id":"2605.22416","citing_title":"Asymmetric Virtual Memory Paging for Hybrid Mamba-Transformer Inference","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2410.17891","citing_title":"Scaling Diffusion Language Models via Adaptation from Autoregressive Models","ref_index":175,"is_internal_anchor":true},{"citing_arxiv_id":"2412.13663","citing_title":"Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference","ref_index":182,"is_internal_anchor":true},{"citing_arxiv_id":"2605.17410","citing_title":"Computational Challenges in Token Economics: Bridging Economic Theory and AI System Design","ref_index":29,"is_internal_anchor":true},{"citing_arxiv_id":"2504.19874","citing_title":"TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate","ref_index":47,"is_internal_anchor":true},{"citing_arxiv_id":"2605.15617","citing_title":"A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM","ref_index":27,"is_internal_anchor":true},{"citing_arxiv_id":"2409.10516","citing_title":"RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2511.00413","citing_title":"Tree Training: Accelerating Agentic LLMs Training via Shared Prefix Reuse","ref_index":13,"is_internal_anchor":true},{"citing_arxiv_id":"2412.05496","citing_title":"Flex Attention: A Programming Model for Generating Optimized Attention Kernels","ref_index":44,"is_internal_anchor":true},{"citing_arxiv_id":"2601.14910","citing_title":"PipeWeave: Synergizing Analytical and Learning Models for Unified GPU Performance Prediction","ref_index":60,"is_internal_anchor":true},{"citing_arxiv_id":"2502.10517","citing_title":"KernelBench: Can LLMs Write Efficient GPU Kernels?","ref_index":30,"is_internal_anchor":true},{"citing_arxiv_id":"2511.18870","citing_title":"HunyuanVideo 1.5 Technical Report","ref_index":7,"is_internal_anchor":true},{"citing_arxiv_id":"2505.13211","citing_title":"MAGI-1: Autoregressive Video Generation at Scale","ref_index":39,"is_internal_anchor":true},{"citing_arxiv_id":"2605.11999","citing_title":"The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.08467","citing_title":"CUDAHercules: Benchmarking Hardware-Aware Expert-level CUDA Optimization for LLMs","ref_index":20,"is_internal_anchor":true},{"citing_arxiv_id":"2605.00555","citing_title":"Sim-FA: A GPGPU Simulator Framework for Fine-Grained FlashAttention Pipeline Analysis","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2604.12163","citing_title":"Nucleus-Image: Sparse MoE for Image Generation","ref_index":23,"is_internal_anchor":true},{"citing_arxiv_id":"2604.22808","citing_title":"FreqFormer: Hierarchical Frequency-Domain Attention with Adaptive Spectral Routing for Long-Sequence Video Diffusion Transformers","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14442","citing_title":"Hierarchical vs. Flat Iteration in Shared-Weight Transformers","ref_index":25,"is_internal_anchor":true},{"citing_arxiv_id":"2604.15379","citing_title":"Fleet: Hierarchical Task-based Abstraction for Megakernels on Multi-Die GPUs","ref_index":6,"is_internal_anchor":true},{"citing_arxiv_id":"2604.14825","citing_title":"Nautilus: An Auto-Scheduling Tensor Compiler for Efficient Tiled GPU Kernels","ref_index":35,"is_internal_anchor":true},{"citing_arxiv_id":"2605.02568","citing_title":"StreamIndex: Memory-Bounded Compressed Sparse Attention via Streaming Top-k","ref_index":27,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG","json":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG.json","graph_json":"https://pith.science/api/pith-number/UQKLUT5OACZ5Y4RG43GMMQZULG/graph.json","events_json":"https://pith.science/api/pith-number/UQKLUT5OACZ5Y4RG43GMMQZULG/events.json","paper":"https://pith.science/paper/UQKLUT5O"},"agent_actions":{"view_html":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG","download_json":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG.json","view_paper":"https://pith.science/paper/UQKLUT5O","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.08608&json=true","fetch_graph":"https://pith.science/api/pith-number/UQKLUT5OACZ5Y4RG43GMMQZULG/graph.json","fetch_events":"https://pith.science/api/pith-number/UQKLUT5OACZ5Y4RG43GMMQZULG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG/action/storage_attestation","attest_author":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG/action/author_attestation","sign_citation":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG/action/citation_signature","submit_replication":"https://pith.science/pith/UQKLUT5OACZ5Y4RG43GMMQZULG/action/replication_record"}},"created_at":"2026-05-20T19:38:40.310648+00:00","updated_at":"2026-05-20T19:38:40.310648+00:00"}