{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:E2CX6AJVRH6RK52ZFHCXM36H5S","short_pith_number":"pith:E2CX6AJV","schema_version":"1.0","canonical_sha256":"26857f013589fd15775929c5766fc7ecbc8bd6b75c77703a9266ed0d605d54f0","source":{"kind":"arxiv","id":"2603.23640","version":2},"attestation_state":"computed","paper":{"title":"LLM Inference at the Edge: Mobile, NPU, and GPU Performance Efficiency Trade-offs Under Sustained Load","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.DC","authors_text":"Kautuk Kundan, Pranay Tummalapalli, Ritam Pal, Sahil Arayakandy","submitted_at":"2026-03-24T18:28:38Z","abstract_excerpt":"Deploying large language models on-device for always-on personal agents demands sustained inference from hardware tightly constrained in power, thermal envelope, and memory. We benchmark Qwen 2.5 1.5B (4-bit quantised) across four platforms: a Raspberry Pi 5 with Hailo-10H NPU, a Samsung Galaxy S24 Ultra, an iPhone 16 Pro, and a laptop NVIDIA RTX 4050 GPU. Using a fixed 258-token prompt over 20 warm-condition iterations per device, we measure throughput, latency, power, and thermal behaviour. For mobile platforms, thermal management supersedes peak compute as the primary constraint: the iPhone"},"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":"2603.23640","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.DC","submitted_at":"2026-03-24T18:28:38Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"26445588945946a372c4c14a73906e38073266127cb50a3c90807310dec06c7e","abstract_canon_sha256":"960f1fbd4a86e2b546afeb63f99bc9c98d6605a10efd759693497ab381a4347d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-09T02:07:24.748635Z","signature_b64":"4IrwIPOQ4/3M+x3USfDlPLs5T46aR8Os17ju9Vbbx0KB/pyO4Hsxu+tsH83ZM/E5FK8atKU44X6/ix9gm+uMCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"26857f013589fd15775929c5766fc7ecbc8bd6b75c77703a9266ed0d605d54f0","last_reissued_at":"2026-06-09T02:07:24.747677Z","signature_status":"signed_v1","first_computed_at":"2026-06-09T02:07:24.747677Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"LLM Inference at the Edge: Mobile, NPU, and GPU Performance Efficiency Trade-offs Under Sustained Load","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.DC","authors_text":"Kautuk Kundan, Pranay Tummalapalli, Ritam Pal, Sahil Arayakandy","submitted_at":"2026-03-24T18:28:38Z","abstract_excerpt":"Deploying large language models on-device for always-on personal agents demands sustained inference from hardware tightly constrained in power, thermal envelope, and memory. We benchmark Qwen 2.5 1.5B (4-bit quantised) across four platforms: a Raspberry Pi 5 with Hailo-10H NPU, a Samsung Galaxy S24 Ultra, an iPhone 16 Pro, and a laptop NVIDIA RTX 4050 GPU. Using a fixed 258-token prompt over 20 warm-condition iterations per device, we measure throughput, latency, power, and thermal behaviour. For mobile platforms, thermal management supersedes peak compute as the primary constraint: the iPhone"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.23640","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/2603.23640/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":"2603.23640","created_at":"2026-06-09T02:07:24.747828+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.23640v2","created_at":"2026-06-09T02:07:24.747828+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.23640","created_at":"2026-06-09T02:07:24.747828+00:00"},{"alias_kind":"pith_short_12","alias_value":"E2CX6AJVRH6R","created_at":"2026-06-09T02:07:24.747828+00:00"},{"alias_kind":"pith_short_16","alias_value":"E2CX6AJVRH6RK52Z","created_at":"2026-06-09T02:07:24.747828+00:00"},{"alias_kind":"pith_short_8","alias_value":"E2CX6AJV","created_at":"2026-06-09T02:07:24.747828+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2604.24785","citing_title":"Cloud to Edge: Benchmarking LLM Inference On Hardware-Accelerated Single-Board Computers","ref_index":18,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S","json":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S.json","graph_json":"https://pith.science/api/pith-number/E2CX6AJVRH6RK52ZFHCXM36H5S/graph.json","events_json":"https://pith.science/api/pith-number/E2CX6AJVRH6RK52ZFHCXM36H5S/events.json","paper":"https://pith.science/paper/E2CX6AJV"},"agent_actions":{"view_html":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S","download_json":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S.json","view_paper":"https://pith.science/paper/E2CX6AJV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.23640&json=true","fetch_graph":"https://pith.science/api/pith-number/E2CX6AJVRH6RK52ZFHCXM36H5S/graph.json","fetch_events":"https://pith.science/api/pith-number/E2CX6AJVRH6RK52ZFHCXM36H5S/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S/action/timestamp_anchor","attest_storage":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S/action/storage_attestation","attest_author":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S/action/author_attestation","sign_citation":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S/action/citation_signature","submit_replication":"https://pith.science/pith/E2CX6AJVRH6RK52ZFHCXM36H5S/action/replication_record"}},"created_at":"2026-06-09T02:07:24.747828+00:00","updated_at":"2026-06-09T02:07:24.747828+00:00"}