{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:LEGVSBB7P6ECT3OCEHTXUUVQRI","short_pith_number":"pith:LEGVSBB7","schema_version":"1.0","canonical_sha256":"590d59043f7f8829edc221e77a52b08a1624f831956896a50a561f0783fe7c1d","source":{"kind":"arxiv","id":"2505.06461","version":1},"attestation_state":"computed","paper":{"title":"Challenging GPU Dominance: When CPUs Outperform for On-Device LLM Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.DC","authors_text":"Haolin Zhang, Jeff Huang","submitted_at":"2025-05-09T23:05:53Z","abstract_excerpt":"The common assumption in on-device AI is that GPUs, with their superior parallel processing, always provide the best performance for large language model (LLM) inference. In this work, we challenge this notion by empirically demonstrating that, under certain conditions, CPUs can outperform GPUs for LLM inference on mobile devices. Using a 1-billion-parameter LLM deployed via llama.cpp on the iPhone 15 Pro, we show that a CPU-only configuration (two threads, F16 precision) achieves 17 tokens per second, surpassing the 12.8 tokens per second obtained with GPU acceleration. We analyze the archite"},"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":"2505.06461","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.DC","submitted_at":"2025-05-09T23:05:53Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"ca7eff2a0a9c122d7c8cd9d3d84687d21d3ddf2de1d1a13fd5e20148c267cc04","abstract_canon_sha256":"da5c4c739f76395c112d76f302ba3d9dbe22d1eb75b4e4ab3da7375c0815a94b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:01:09.991583Z","signature_b64":"a8uMewDAGa2lpGC/Fzan10Yamx6WXdOsm2mCLZrHXS7VByGcbEv2ncohCb8JfqYtyZXWbqI6S7NAkUWjTLuWBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"590d59043f7f8829edc221e77a52b08a1624f831956896a50a561f0783fe7c1d","last_reissued_at":"2026-07-05T11:01:09.991104Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:01:09.991104Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Challenging GPU Dominance: When CPUs Outperform for On-Device LLM Inference","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.DC","authors_text":"Haolin Zhang, Jeff Huang","submitted_at":"2025-05-09T23:05:53Z","abstract_excerpt":"The common assumption in on-device AI is that GPUs, with their superior parallel processing, always provide the best performance for large language model (LLM) inference. In this work, we challenge this notion by empirically demonstrating that, under certain conditions, CPUs can outperform GPUs for LLM inference on mobile devices. Using a 1-billion-parameter LLM deployed via llama.cpp on the iPhone 15 Pro, we show that a CPU-only configuration (two threads, F16 precision) achieves 17 tokens per second, surpassing the 12.8 tokens per second obtained with GPU acceleration. We analyze the archite"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.06461","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/2505.06461/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":"2505.06461","created_at":"2026-07-05T11:01:09.991160+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.06461v1","created_at":"2026-07-05T11:01:09.991160+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.06461","created_at":"2026-07-05T11:01:09.991160+00:00"},{"alias_kind":"pith_short_12","alias_value":"LEGVSBB7P6EC","created_at":"2026-07-05T11:01:09.991160+00:00"},{"alias_kind":"pith_short_16","alias_value":"LEGVSBB7P6ECT3OC","created_at":"2026-07-05T11:01:09.991160+00:00"},{"alias_kind":"pith_short_8","alias_value":"LEGVSBB7","created_at":"2026-07-05T11:01:09.991160+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.27435","citing_title":"When NPUs Are Not Always Faster: A Stage-Level Analysis of Mobile LLM Inference","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI","json":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI.json","graph_json":"https://pith.science/api/pith-number/LEGVSBB7P6ECT3OCEHTXUUVQRI/graph.json","events_json":"https://pith.science/api/pith-number/LEGVSBB7P6ECT3OCEHTXUUVQRI/events.json","paper":"https://pith.science/paper/LEGVSBB7"},"agent_actions":{"view_html":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI","download_json":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI.json","view_paper":"https://pith.science/paper/LEGVSBB7","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.06461&json=true","fetch_graph":"https://pith.science/api/pith-number/LEGVSBB7P6ECT3OCEHTXUUVQRI/graph.json","fetch_events":"https://pith.science/api/pith-number/LEGVSBB7P6ECT3OCEHTXUUVQRI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI/action/storage_attestation","attest_author":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI/action/author_attestation","sign_citation":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI/action/citation_signature","submit_replication":"https://pith.science/pith/LEGVSBB7P6ECT3OCEHTXUUVQRI/action/replication_record"}},"created_at":"2026-07-05T11:01:09.991160+00:00","updated_at":"2026-07-05T11:01:09.991160+00:00"}