{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2020:4FSP6Q6IVQLZFNZLAJZ6H5WXVF","short_pith_number":"pith:4FSP6Q6I","schema_version":"1.0","canonical_sha256":"e164ff43c8ac1792b72b0273e3f6d7a94fa3a77619430fac684e27b869d49b59","source":{"kind":"arxiv","id":"2010.06192","version":2},"attestation_state":"computed","paper":{"title":"Revisiting BFloat16 Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["stat.ML"],"primary_cat":"cs.LG","authors_text":"Christopher De Sa, Christopher R. Aberger, Jian Zhang, Pedram Zamirai","submitted_at":"2020-10-13T05:38:07Z","abstract_excerpt":"State-of-the-art generic low-precision training algorithms use a mix of 16-bit and 32-bit precision, creating the folklore that 16-bit hardware compute units alone are not enough to maximize model accuracy. As a result, deep learning accelerators are forced to support both 16-bit and 32-bit floating-point units (FPUs), which is more costly than only using 16-bit FPUs for hardware design. We ask: can we train deep learning models only with 16-bit floating-point units, while still matching the model accuracy attained by 32-bit training? Towards this end, we study 16-bit-FPU training on the widel"},"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":"2010.06192","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2020-10-13T05:38:07Z","cross_cats_sorted":["stat.ML"],"title_canon_sha256":"8543c9295ec23a6bf9208274f42ff2317005ea5d5a45061273408a597ad6f020","abstract_canon_sha256":"b076fe61783f0fccb616ad0f262bde9f9252f8df3a72318a606dbccab2923202"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T02:20:57.197484Z","signature_b64":"mUiuyOPFYdtN63WoK88GrZT4dDzl0Cwmty+ztI3QrldfVIibxYPm1kX+Q8IhI/vtHIi9pwIb1ermFwRtOM8ODQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e164ff43c8ac1792b72b0273e3f6d7a94fa3a77619430fac684e27b869d49b59","last_reissued_at":"2026-07-05T02:20:57.197039Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T02:20:57.197039Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Revisiting BFloat16 Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["stat.ML"],"primary_cat":"cs.LG","authors_text":"Christopher De Sa, Christopher R. Aberger, Jian Zhang, Pedram Zamirai","submitted_at":"2020-10-13T05:38:07Z","abstract_excerpt":"State-of-the-art generic low-precision training algorithms use a mix of 16-bit and 32-bit precision, creating the folklore that 16-bit hardware compute units alone are not enough to maximize model accuracy. As a result, deep learning accelerators are forced to support both 16-bit and 32-bit floating-point units (FPUs), which is more costly than only using 16-bit FPUs for hardware design. We ask: can we train deep learning models only with 16-bit floating-point units, while still matching the model accuracy attained by 32-bit training? Towards this end, we study 16-bit-FPU training on the widel"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2010.06192","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/2010.06192/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":"2010.06192","created_at":"2026-07-05T02:20:57.197098+00:00"},{"alias_kind":"arxiv_version","alias_value":"2010.06192v2","created_at":"2026-07-05T02:20:57.197098+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2010.06192","created_at":"2026-07-05T02:20:57.197098+00:00"},{"alias_kind":"pith_short_12","alias_value":"4FSP6Q6IVQLZ","created_at":"2026-07-05T02:20:57.197098+00:00"},{"alias_kind":"pith_short_16","alias_value":"4FSP6Q6IVQLZFNZL","created_at":"2026-07-05T02:20:57.197098+00:00"},{"alias_kind":"pith_short_8","alias_value":"4FSP6Q6I","created_at":"2026-07-05T02:20:57.197098+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2502.05171","citing_title":"Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach","ref_index":177,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF","json":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF.json","graph_json":"https://pith.science/api/pith-number/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/graph.json","events_json":"https://pith.science/api/pith-number/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/events.json","paper":"https://pith.science/paper/4FSP6Q6I"},"agent_actions":{"view_html":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF","download_json":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF.json","view_paper":"https://pith.science/paper/4FSP6Q6I","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2010.06192&json=true","fetch_graph":"https://pith.science/api/pith-number/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/graph.json","fetch_events":"https://pith.science/api/pith-number/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/action/storage_attestation","attest_author":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/action/author_attestation","sign_citation":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/action/citation_signature","submit_replication":"https://pith.science/pith/4FSP6Q6IVQLZFNZLAJZ6H5WXVF/action/replication_record"}},"created_at":"2026-07-05T02:20:57.197098+00:00","updated_at":"2026-07-05T02:20:57.197098+00:00"}