{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2025:LEPUH3LRFIXTB5OR22JYS7OHJQ","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"1c276ad2095a7b3e426b7bf9defb888ab9216db718737a0b3515565e9a1c3757","cross_cats_sorted":["cs.NI","cs.PF"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.DC","submitted_at":"2025-12-18T00:45:00Z","title_canon_sha256":"bee3c6b1a798f883f47117bc5ddb7809fff2573bae5d05669f3b5ee9bca3ad8c"},"schema_version":"1.0","source":{"id":"2512.16056","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2512.16056","created_at":"2026-05-18T02:44:32Z"},{"alias_kind":"arxiv_version","alias_value":"2512.16056v2","created_at":"2026-05-18T02:44:32Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2512.16056","created_at":"2026-05-18T02:44:32Z"},{"alias_kind":"pith_short_12","alias_value":"LEPUH3LRFIXT","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"LEPUH3LRFIXTB5OR","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"LEPUH3LR","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:d11fdd92a2749a6f96f1f54e7a5a2fb244c3a726cfd47bbb74e69e3f141b702f","target":"graph","created_at":"2026-05-18T02:44:32Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"On an 8-GPU NVIDIA H20 server, MMA achieves 245 GB/s peak host-to-GPU bandwidth, a 4.62x improvement over native CUDA copies, and reduces TTFT for KV cache fetching by 1.14-2.38x and model wake-up/switching latency by 1.12-2.48x."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the additional relay paths through peer GPUs and high-bandwidth interconnects can be utilized with low enough overhead and without breaking CUDA stream semantics or application correctness under real LLM workloads."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"MMA routes host-GPU transfers over multiple available paths to deliver 4.62x higher peak bandwidth and lower latencies in LLM serving without hardware or driver changes."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Multipath Memory Access routes host-GPU copies over unused server links to raise bandwidth 4.6x."}],"snapshot_sha256":"7a9b940f8b6997b6775ec9caeea74b659146f4f8fdfb5dfdef8a46921900e8ff"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"887e95af28021118d353108b58ceb78b080a7a286f5a7d98d9413c2e06346528"},"paper":{"abstract_excerpt":"Host-GPU data movement has become a latency-critical bottleneck in LLM serving, surfacing in common paths such as model-weight movement and KV cache offload/fetch. Today, each host-GPU copy is effectively confined to the PCIe path of the target GPU, even though modern multi-GPU servers contain additional PCIe links on peer GPUs and high bandwidth GPU interconnects. This leaves substantial intra-server I/O capacity unused. To address this issue, we present Multipath Memory Access (MMA), a software-defined multipath memory access system for host--GPU data transfer. To the best of our knowledge, ","authors_text":"Chengguang Xu, Daoping Zhang, Feiqiang Sun, Feng Jin, Guo Chen, Junjie Chen, Lingfeng Tang, Peihao Huang, Yuxin Chen","cross_cats":["cs.NI","cs.PF"],"headline":"Multipath Memory Access routes host-GPU copies over unused server links to raise bandwidth 4.6x.","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.DC","submitted_at":"2025-12-18T00:45:00Z","title":"MultiPath Memory Access: Breaking Host-GPU Bandwidth Bottlenecks in LLM Services"},"references":{"count":52,"internal_anchors":3,"resolved_work":52,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"MMA code base will be opensourced after paper ac- cepted","work_id":"62fb66e6-930b-43a1-be90-57ee36b0ba4a","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"Advanced Micro Devices","work_id":"9be9b912-df33-4373-a31f-11476d1f55b2","year":2022},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":3,"title":"Advanced Micro Devices","work_id":"89774536-8edc-4183-ac14-ed01cf54940b","year":2024},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Deepspeed-inference: enabling efficient in- ference of transformer models at unprecedented scale","work_id":"461bf7aa-e1b9-4309-a540-e2905d823c31","year":2022},{"cited_arxiv_id":"2309.16609","doi":"","is_internal_anchor":true,"ref_index":5,"title":"Qwen Technical Report","work_id":"bb1fd52f-6b2f-437c-9516-37bdf6eb9be8","year":2023}],"snapshot_sha256":"41433ff6ca49d50f8face4bafba8465b901f73a62d0613b2a08589f02906c2a9"},"source":{"id":"2512.16056","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-16T21:54:28.884047Z","id":"aa1d4163-20f0-4562-a277-c1fdf22739c5","model_set":{"reader":"grok-4.3"},"one_line_summary":"MMA routes host-GPU transfers over multiple available paths to deliver 4.62x higher peak bandwidth and lower latencies in LLM serving without hardware or driver changes.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Multipath Memory Access routes host-GPU copies over unused server links to raise bandwidth 4.6x.","strongest_claim":"On an 8-GPU NVIDIA H20 server, MMA achieves 245 GB/s peak host-to-GPU bandwidth, a 4.62x improvement over native CUDA copies, and reduces TTFT for KV cache fetching by 1.14-2.38x and model wake-up/switching latency by 1.12-2.48x.","weakest_assumption":"That the additional relay paths through peer GPUs and high-bandwidth interconnects can be utilized with low enough overhead and without breaking CUDA stream semantics or application correctness under real LLM workloads."}},"verdict_id":"aa1d4163-20f0-4562-a277-c1fdf22739c5"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:8751a4f565edfda3a71ccc462895b4a3627e74a4b7a811b5f9cab9b8c4dfbe36","target":"record","created_at":"2026-05-18T02:44:32Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"1c276ad2095a7b3e426b7bf9defb888ab9216db718737a0b3515565e9a1c3757","cross_cats_sorted":["cs.NI","cs.PF"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.DC","submitted_at":"2025-12-18T00:45:00Z","title_canon_sha256":"bee3c6b1a798f883f47117bc5ddb7809fff2573bae5d05669f3b5ee9bca3ad8c"},"schema_version":"1.0","source":{"id":"2512.16056","kind":"arxiv","version":2}},"canonical_sha256":"591f43ed712a2f30f5d1d693897dc74c245222d973015ad300b790f68fc4b306","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"591f43ed712a2f30f5d1d693897dc74c245222d973015ad300b790f68fc4b306","first_computed_at":"2026-05-18T02:44:32.141856Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T02:44:32.141856Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"eR1GunCgyG48Ik8to4qCzaMebyhYJy9plyCUfBH6x9wVHDP8Ex3+Tw+Bg9i4NUmHSdb5gGYrGqm3R3QB7BXLAA==","signature_status":"signed_v1","signed_at":"2026-05-18T02:44:32.142287Z","signed_message":"canonical_sha256_bytes"},"source_id":"2512.16056","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:8751a4f565edfda3a71ccc462895b4a3627e74a4b7a811b5f9cab9b8c4dfbe36","sha256:d11fdd92a2749a6f96f1f54e7a5a2fb244c3a726cfd47bbb74e69e3f141b702f"],"state_sha256":"b8138b86e8b58d3ee175dc4f52d9453e21b1d6acfbaac5f361b099248b671f80"}