{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:ENF2DUYOJSHMMBJPR5E4LGEK5L","short_pith_number":"pith:ENF2DUYO","schema_version":"1.0","canonical_sha256":"234ba1d30e4c8ec6052f8f49c5988aeaea26dc18b1432a8f12efdf1ec09f8987","source":{"kind":"arxiv","id":"2504.17584","version":1},"attestation_state":"computed","paper":{"title":"L3: DIMM-PIM Integrated Architecture and Coordination for Scalable Long-Context LLM Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AR","authors_text":"Dong Du, Haibo Chen, Haocheng Wang, Liyan Chen, Naifeng Jing, Qingyuan Liu, Yanning Yang, Yubin Xia, Zhigang Mao","submitted_at":"2025-04-24T14:14:07Z","abstract_excerpt":"Large Language Models (LLMs) increasingly require processing long text sequences, but GPU memory limitations force difficult trade-offs between memory capacity and bandwidth. While HBM-based acceleration offers high bandwidth, its capacity remains constrained. Offloading data to host-side DIMMs improves capacity but introduces costly data swapping overhead. We identify that the critical memory bottleneck lies in the decoding phase of multi-head attention (MHA) exclusively, which demands substantial capacity for storing KV caches and high bandwidth for attention computation. Our key insight rev"},"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":"2504.17584","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AR","submitted_at":"2025-04-24T14:14:07Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"96fec74f4adbe754e49f16dd9e581d4c5fef0a75cb62d4d7e99e6cd47aadd1e1","abstract_canon_sha256":"c3b6c0a96063c7e91d6a01767740e14a398257737a30e52c41bc93a739abe34c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:53:31.702377Z","signature_b64":"hoMn6ag31YeywnQi9DOVZsI1iDDO+VDHbPphfdiX7/X0SbJU232CcSz+UBe5QqirDwheD74NBdp/BIP57G5ICQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"234ba1d30e4c8ec6052f8f49c5988aeaea26dc18b1432a8f12efdf1ec09f8987","last_reissued_at":"2026-07-05T10:53:31.701869Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:53:31.701869Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"L3: DIMM-PIM Integrated Architecture and Coordination for Scalable Long-Context LLM Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AR","authors_text":"Dong Du, Haibo Chen, Haocheng Wang, Liyan Chen, Naifeng Jing, Qingyuan Liu, Yanning Yang, Yubin Xia, Zhigang Mao","submitted_at":"2025-04-24T14:14:07Z","abstract_excerpt":"Large Language Models (LLMs) increasingly require processing long text sequences, but GPU memory limitations force difficult trade-offs between memory capacity and bandwidth. While HBM-based acceleration offers high bandwidth, its capacity remains constrained. Offloading data to host-side DIMMs improves capacity but introduces costly data swapping overhead. We identify that the critical memory bottleneck lies in the decoding phase of multi-head attention (MHA) exclusively, which demands substantial capacity for storing KV caches and high bandwidth for attention computation. Our key insight rev"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.17584","kind":"arxiv","version":1},"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/2504.17584/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":"2504.17584","created_at":"2026-07-05T10:53:31.701937+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.17584v1","created_at":"2026-07-05T10:53:31.701937+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.17584","created_at":"2026-07-05T10:53:31.701937+00:00"},{"alias_kind":"pith_short_12","alias_value":"ENF2DUYOJSHM","created_at":"2026-07-05T10:53:31.701937+00:00"},{"alias_kind":"pith_short_16","alias_value":"ENF2DUYOJSHMMBJP","created_at":"2026-07-05T10:53:31.701937+00:00"},{"alias_kind":"pith_short_8","alias_value":"ENF2DUYO","created_at":"2026-07-05T10:53:31.701937+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.05639","citing_title":"TokenStack: A Heterogeneous HBM-PIM Architecture and Runtime for Efficient LLM Inference","ref_index":35,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L","json":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L.json","graph_json":"https://pith.science/api/pith-number/ENF2DUYOJSHMMBJPR5E4LGEK5L/graph.json","events_json":"https://pith.science/api/pith-number/ENF2DUYOJSHMMBJPR5E4LGEK5L/events.json","paper":"https://pith.science/paper/ENF2DUYO"},"agent_actions":{"view_html":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L","download_json":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L.json","view_paper":"https://pith.science/paper/ENF2DUYO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.17584&json=true","fetch_graph":"https://pith.science/api/pith-number/ENF2DUYOJSHMMBJPR5E4LGEK5L/graph.json","fetch_events":"https://pith.science/api/pith-number/ENF2DUYOJSHMMBJPR5E4LGEK5L/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L/action/storage_attestation","attest_author":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L/action/author_attestation","sign_citation":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L/action/citation_signature","submit_replication":"https://pith.science/pith/ENF2DUYOJSHMMBJPR5E4LGEK5L/action/replication_record"}},"created_at":"2026-07-05T10:53:31.701937+00:00","updated_at":"2026-07-05T10:53:31.701937+00:00"}