{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:Z7TPZI636O5ZOQCTSY7RB6FXP7","short_pith_number":"pith:Z7TPZI63","schema_version":"1.0","canonical_sha256":"cfe6fca3dbf3bb974053963f10f8b77fd89d7633fed379f0e77d497353f568ec","source":{"kind":"arxiv","id":"2408.12757","version":2},"attestation_state":"computed","paper":{"title":"NanoFlow: Towards Optimal Large Language Model Serving Throughput","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.DC","authors_text":"Arvind Krishnamurthy, Baris Kasikci, Chien-Yu Lin, Dedong Xie, Gefei Zuo, Kan Zhu, Keisuke Kamahori, Liangyu Zhao, Qinyu Xu, Stephanie Wang, Tian Tang, Yile Gu, Yilong Zhao, Yufei Gao, Zihao Ye, Ziren Wang","submitted_at":"2024-08-22T23:00:40Z","abstract_excerpt":"Large Language Models (LLMs) have resulted in a surging demand for planet-scale serving systems, where tens of thousands of GPUs continuously serve hundreds of millions of users. Consequently, throughput has emerged as a key metric that determines serving systems' performance. Due to large model sizes and memory-intensive self-attention, LLM serving has been commonly assumed to be memory-bound. Through a detailed analysis, we show that despite having memory-intensive components, end-to-end LLM serving is compute bound for most common workloads and LLMs. Alas, most existing serving engines fall"},"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":"2408.12757","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.DC","submitted_at":"2024-08-22T23:00:40Z","cross_cats_sorted":[],"title_canon_sha256":"f0dd478d1c47e335eea933aded09020cb48ea72d19c3611eeb1cd10739ceb9df","abstract_canon_sha256":"f4ec8e77cd5d9922d1d63fadfc6c777b9572c2d21d5ea53ea71f12cd889bf479"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:08:59.563122Z","signature_b64":"2SD0VonSjXuiJeHmQ9KA5w5/0mJJWDf922AT10DbmRCpDdot9fPXoVM+XURNF/Y+E0mTi1EG/5Iw+xuXIZdWDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cfe6fca3dbf3bb974053963f10f8b77fd89d7633fed379f0e77d497353f568ec","last_reissued_at":"2026-07-05T11:08:59.562602Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:08:59.562602Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"NanoFlow: Towards Optimal Large Language Model Serving Throughput","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.DC","authors_text":"Arvind Krishnamurthy, Baris Kasikci, Chien-Yu Lin, Dedong Xie, Gefei Zuo, Kan Zhu, Keisuke Kamahori, Liangyu Zhao, Qinyu Xu, Stephanie Wang, Tian Tang, Yile Gu, Yilong Zhao, Yufei Gao, Zihao Ye, Ziren Wang","submitted_at":"2024-08-22T23:00:40Z","abstract_excerpt":"Large Language Models (LLMs) have resulted in a surging demand for planet-scale serving systems, where tens of thousands of GPUs continuously serve hundreds of millions of users. Consequently, throughput has emerged as a key metric that determines serving systems' performance. Due to large model sizes and memory-intensive self-attention, LLM serving has been commonly assumed to be memory-bound. Through a detailed analysis, we show that despite having memory-intensive components, end-to-end LLM serving is compute bound for most common workloads and LLMs. Alas, most existing serving engines fall"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2408.12757","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/2408.12757/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":"2408.12757","created_at":"2026-07-05T11:08:59.562666+00:00"},{"alias_kind":"arxiv_version","alias_value":"2408.12757v2","created_at":"2026-07-05T11:08:59.562666+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2408.12757","created_at":"2026-07-05T11:08:59.562666+00:00"},{"alias_kind":"pith_short_12","alias_value":"Z7TPZI636O5Z","created_at":"2026-07-05T11:08:59.562666+00:00"},{"alias_kind":"pith_short_16","alias_value":"Z7TPZI636O5ZOQCT","created_at":"2026-07-05T11:08:59.562666+00:00"},{"alias_kind":"pith_short_8","alias_value":"Z7TPZI63","created_at":"2026-07-05T11:08:59.562666+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.20295","citing_title":"Token-Operations-Oriented Inference Optimization Techniques for Large Models","ref_index":211,"is_internal_anchor":false},{"citing_arxiv_id":"2606.19746","citing_title":"SAC: Disaggregated KV Cache System for Sparse Attention LLMs with CXL","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10493","citing_title":"Achieving Cloud-Grade SLOs for Local Mixture-of-Experts Inference through CPU-GPU Hybrid Design","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2606.09200","citing_title":"Resource-aware Computation-Communication Overlap for multi-GPU ML Workloads","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24203","citing_title":"Agentic Witnessing: Pragmatic and Scalable TEE-Enabled Privacy-Preserving Auditing","ref_index":45,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7","json":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7.json","graph_json":"https://pith.science/api/pith-number/Z7TPZI636O5ZOQCTSY7RB6FXP7/graph.json","events_json":"https://pith.science/api/pith-number/Z7TPZI636O5ZOQCTSY7RB6FXP7/events.json","paper":"https://pith.science/paper/Z7TPZI63"},"agent_actions":{"view_html":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7","download_json":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7.json","view_paper":"https://pith.science/paper/Z7TPZI63","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2408.12757&json=true","fetch_graph":"https://pith.science/api/pith-number/Z7TPZI636O5ZOQCTSY7RB6FXP7/graph.json","fetch_events":"https://pith.science/api/pith-number/Z7TPZI636O5ZOQCTSY7RB6FXP7/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7/action/timestamp_anchor","attest_storage":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7/action/storage_attestation","attest_author":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7/action/author_attestation","sign_citation":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7/action/citation_signature","submit_replication":"https://pith.science/pith/Z7TPZI636O5ZOQCTSY7RB6FXP7/action/replication_record"}},"created_at":"2026-07-05T11:08:59.562666+00:00","updated_at":"2026-07-05T11:08:59.562666+00:00"}