{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2021:2PGW4AHIK4E5LRPHQQGNLVDANR","short_pith_number":"pith:2PGW4AHI","canonical_record":{"source":{"id":"2106.08295","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2021-06-15T17:12:42Z","cross_cats_sorted":["cs.AI","cs.CV"],"title_canon_sha256":"d400ea5a0f749d08b93527e85edb6bcd209e23bedae391ebe6b390bafcff6167","abstract_canon_sha256":"6653bc600ac410686a82af6846d19281d6384e8032eab643122276bdf6c6abae"},"schema_version":"1.0"},"canonical_sha256":"d3cd6e00e85709d5c5e7840cd5d4606c5a81575a9ce8683db98743c571d9d1a6","source":{"kind":"arxiv","id":"2106.08295","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2106.08295","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"arxiv_version","alias_value":"2106.08295v1","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2106.08295","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"pith_short_12","alias_value":"2PGW4AHIK4E5","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"2PGW4AHIK4E5LRPH","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"2PGW4AHI","created_at":"2026-05-18T12:33:33Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2021:2PGW4AHIK4E5LRPHQQGNLVDANR","target":"record","payload":{"canonical_record":{"source":{"id":"2106.08295","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2021-06-15T17:12:42Z","cross_cats_sorted":["cs.AI","cs.CV"],"title_canon_sha256":"d400ea5a0f749d08b93527e85edb6bcd209e23bedae391ebe6b390bafcff6167","abstract_canon_sha256":"6653bc600ac410686a82af6846d19281d6384e8032eab643122276bdf6c6abae"},"schema_version":"1.0"},"canonical_sha256":"d3cd6e00e85709d5c5e7840cd5d4606c5a81575a9ce8683db98743c571d9d1a6","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:15.165185Z","signature_b64":"XDgLEuRr28O94zAEZzTM9IvL4JMMxLZqMfExpz6F/kiq231cG+eFqo9btpTQ38okTYfom2+lLRENQFi/JJ3SDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d3cd6e00e85709d5c5e7840cd5d4606c5a81575a9ce8683db98743c571d9d1a6","last_reissued_at":"2026-05-17T23:38:15.164398Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:15.164398Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2106.08295","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"YlNAXnO6Wu3X90k4Pojfv0HDO+wRry8kJea2CB4tpljxvFbRypqFMJiPIGQfbQcxuEF77CbRUMc6Lm6girNIDw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-29T16:11:38.963262Z"},"content_sha256":"45de974b6886d3fa548cca162761bc4fd5e1196084d7a243edf42feff66b0fc9","schema_version":"1.0","event_id":"sha256:45de974b6886d3fa548cca162761bc4fd5e1196084d7a243edf42feff66b0fc9"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2021:2PGW4AHIK4E5LRPHQQGNLVDANR","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"A White Paper on Neural Network Quantization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Post-training quantization suffices for 8-bit neural networks with near floating-point accuracy in most cases.","cross_cats":["cs.AI","cs.CV"],"primary_cat":"cs.LG","authors_text":"Marios Fournarakis, Markus Nagel, Mart van Baalen, Rana Ali Amjad, Tijmen Blankevoort, Yelysei Bondarenko","submitted_at":"2021-06-15T17:12:42Z","abstract_excerpt":"While neural networks have advanced the frontiers in many applications, they often come at a high computational cost. Reducing the power and latency of neural network inference is key if we want to integrate modern networks into edge devices with strict power and compute requirements. Neural network quantization is one of the most effective ways of achieving these savings but the additional noise it induces can lead to accuracy degradation. In this white paper, we introduce state-of-the-art algorithms for mitigating the impact of quantization noise on the network's performance while maintainin"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the described pipelines and state-of-the-art algorithms will deliver the claimed accuracy on arbitrary models and tasks without further model-specific adjustments or access to representative data.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Provides tested PTQ and QAT pipelines that achieve near floating-point accuracy at 8 bits and competitive results at lower bits for common deep learning models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Post-training quantization suffices for 8-bit neural networks with near floating-point accuracy in most cases.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d520db5fc57c86d2b8515b668f73b1c8b08be74eb553c78a1f014455d005a27d"},"source":{"id":"2106.08295","kind":"arxiv","version":1},"verdict":{"id":"a67f9bc1-9111-43a4-b722-cddcef211c48","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T04:06:35.852770Z","strongest_claim":"In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results.","one_line_summary":"Provides tested PTQ and QAT pipelines that achieve near floating-point accuracy at 8 bits and competitive results at lower bits for common deep learning models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the described pipelines and state-of-the-art algorithms will deliver the claimed accuracy on arbitrary models and tasks without further model-specific adjustments or access to representative data.","pith_extraction_headline":"Post-training quantization suffices for 8-bit neural networks with near floating-point accuracy in most cases."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"07bccb3ac47f4ea2adaa6dbe4838bf34772237817bb69684bc94cdcc8876af34"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"a67f9bc1-9111-43a4-b722-cddcef211c48"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"07rpKt89hqxcRQeS0NlSNKXN/i9Mh0GrfMGnsqxnkHkB1JMv6Cbk4d6/0bQnV+loYOGwpLZldAIFGdU1MwUhDw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-29T16:11:38.964171Z"},"content_sha256":"d8dfb5666f78d510165bb6af65e009dd754fec26ba20cc475d0945586e101cc4","schema_version":"1.0","event_id":"sha256:d8dfb5666f78d510165bb6af65e009dd754fec26ba20cc475d0945586e101cc4"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/2PGW4AHIK4E5LRPHQQGNLVDANR/bundle.json","state_url":"https://pith.science/pith/2PGW4AHIK4E5LRPHQQGNLVDANR/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/2PGW4AHIK4E5LRPHQQGNLVDANR/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-29T16:11:38Z","links":{"resolver":"https://pith.science/pith/2PGW4AHIK4E5LRPHQQGNLVDANR","bundle":"https://pith.science/pith/2PGW4AHIK4E5LRPHQQGNLVDANR/bundle.json","state":"https://pith.science/pith/2PGW4AHIK4E5LRPHQQGNLVDANR/state.json","well_known_bundle":"https://pith.science/.well-known/pith/2PGW4AHIK4E5LRPHQQGNLVDANR/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2021:2PGW4AHIK4E5LRPHQQGNLVDANR","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":"6653bc600ac410686a82af6846d19281d6384e8032eab643122276bdf6c6abae","cross_cats_sorted":["cs.AI","cs.CV"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2021-06-15T17:12:42Z","title_canon_sha256":"d400ea5a0f749d08b93527e85edb6bcd209e23bedae391ebe6b390bafcff6167"},"schema_version":"1.0","source":{"id":"2106.08295","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2106.08295","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"arxiv_version","alias_value":"2106.08295v1","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2106.08295","created_at":"2026-05-17T23:38:15Z"},{"alias_kind":"pith_short_12","alias_value":"2PGW4AHIK4E5","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_16","alias_value":"2PGW4AHIK4E5LRPH","created_at":"2026-05-18T12:33:33Z"},{"alias_kind":"pith_short_8","alias_value":"2PGW4AHI","created_at":"2026-05-18T12:33:33Z"}],"graph_snapshots":[{"event_id":"sha256:d8dfb5666f78d510165bb6af65e009dd754fec26ba20cc475d0945586e101cc4","target":"graph","created_at":"2026-05-17T23:38:15Z","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":"In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the described pipelines and state-of-the-art algorithms will deliver the claimed accuracy on arbitrary models and tasks without further model-specific adjustments or access to representative data."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Provides tested PTQ and QAT pipelines that achieve near floating-point accuracy at 8 bits and competitive results at lower bits for common deep learning models."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Post-training quantization suffices for 8-bit neural networks with near floating-point accuracy in most cases."}],"snapshot_sha256":"d520db5fc57c86d2b8515b668f73b1c8b08be74eb553c78a1f014455d005a27d"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"07bccb3ac47f4ea2adaa6dbe4838bf34772237817bb69684bc94cdcc8876af34"},"paper":{"abstract_excerpt":"While neural networks have advanced the frontiers in many applications, they often come at a high computational cost. Reducing the power and latency of neural network inference is key if we want to integrate modern networks into edge devices with strict power and compute requirements. Neural network quantization is one of the most effective ways of achieving these savings but the additional noise it induces can lead to accuracy degradation. In this white paper, we introduce state-of-the-art algorithms for mitigating the impact of quantization noise on the network's performance while maintainin","authors_text":"Marios Fournarakis, Markus Nagel, Mart van Baalen, Rana Ali Amjad, Tijmen Blankevoort, Yelysei Bondarenko","cross_cats":["cs.AI","cs.CV"],"headline":"Post-training quantization suffices for 8-bit neural networks with near floating-point accuracy in most cases.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2021-06-15T17:12:42Z","title":"A White Paper on Neural Network Quantization"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2106.08295","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-17T04:06:35.852770Z","id":"a67f9bc1-9111-43a4-b722-cddcef211c48","model_set":{"reader":"grok-4.3"},"one_line_summary":"Provides tested PTQ and QAT pipelines that achieve near floating-point accuracy at 8 bits and competitive results at lower bits for common deep learning models.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Post-training quantization suffices for 8-bit neural networks with near floating-point accuracy in most cases.","strongest_claim":"In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results.","weakest_assumption":"That the described pipelines and state-of-the-art algorithms will deliver the claimed accuracy on arbitrary models and tasks without further model-specific adjustments or access to representative data."}},"verdict_id":"a67f9bc1-9111-43a4-b722-cddcef211c48"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:45de974b6886d3fa548cca162761bc4fd5e1196084d7a243edf42feff66b0fc9","target":"record","created_at":"2026-05-17T23:38:15Z","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":"6653bc600ac410686a82af6846d19281d6384e8032eab643122276bdf6c6abae","cross_cats_sorted":["cs.AI","cs.CV"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2021-06-15T17:12:42Z","title_canon_sha256":"d400ea5a0f749d08b93527e85edb6bcd209e23bedae391ebe6b390bafcff6167"},"schema_version":"1.0","source":{"id":"2106.08295","kind":"arxiv","version":1}},"canonical_sha256":"d3cd6e00e85709d5c5e7840cd5d4606c5a81575a9ce8683db98743c571d9d1a6","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"d3cd6e00e85709d5c5e7840cd5d4606c5a81575a9ce8683db98743c571d9d1a6","first_computed_at":"2026-05-17T23:38:15.164398Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:15.164398Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"XDgLEuRr28O94zAEZzTM9IvL4JMMxLZqMfExpz6F/kiq231cG+eFqo9btpTQ38okTYfom2+lLRENQFi/JJ3SDQ==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:15.165185Z","signed_message":"canonical_sha256_bytes"},"source_id":"2106.08295","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:45de974b6886d3fa548cca162761bc4fd5e1196084d7a243edf42feff66b0fc9","sha256:d8dfb5666f78d510165bb6af65e009dd754fec26ba20cc475d0945586e101cc4"],"state_sha256":"29f6b31f519030435b0895d1cba655a121409aa684e0e7f4b4c4627b2a917385"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"bTqKXAs0G+8jlWIMHw/in7zuzC8349hGlA1Yjn6i9KKBCfZXgOUoYisRC2+iW8G/cyxibdbdT6KqJUQDp+AvCQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-29T16:11:38.968309Z","bundle_sha256":"8c31a34a3cf99a8669f5b4cdc96052cee79546a76a60ece74c004e70d77725fd"}}