{"paper":{"title":"FP8 Formats for Deep Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"FP8 with E4M3 and E5M2 encodings matches 16-bit training accuracy on large language and image models without hyperparameter changes.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Alexander Heinecke, Dusan Stosic, Hao Wu, John Kamalu, Marius Cornea, Michael Siu, Mohammad Shoeybi, Naveen Mellempudi, Neil Burgess, Patrick Judd, Paulius Micikevicius, Pradeep Dubey, Richard Grisenthwaite, Sangwon Ha, Stuart Oberman","submitted_at":"2022-09-12T17:39:55Z","abstract_excerpt":"FP8 is a natural progression for accelerating deep learning training inference beyond the 16-bit formats common in modern processors. In this paper we propose an 8-bit floating point (FP8) binary interchange format consisting of two encodings - E4M3 (4-bit exponent and 3-bit mantissa) and E5M2 (5-bit exponent and 2-bit mantissa). While E5M2 follows IEEE 754 conventions for representatio of special values, E4M3's dynamic range is extended by not representing infinities and having only one mantissa bit-pattern for NaNs. We demonstrate the efficacy of the FP8 format on a variety of image and lang"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate the efficacy of the FP8 format on a variety of image and language tasks, effectively matching the result quality achieved by 16-bit training sessions. Our study covers the main modern neural network architectures - CNNs, RNNs, and Transformer-based models, leaving all the hyperparameters unchanged from the 16-bit baseline training sessions. Our training experiments include large, up to 175B parameter, language models.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen E4M3 and E5M2 encodings will preserve accuracy across all tasks and model scales without any hyperparameter retuning or task-specific adjustments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FP8 formats E4M3 and E5M2 match 16-bit training accuracy on CNNs, RNNs, and Transformers up to 175B parameters without hyperparameter changes.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FP8 with E4M3 and E5M2 encodings matches 16-bit training accuracy on large language and image models without hyperparameter changes.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7bbd391a1fd0fbfb0c7a1b5b1343c74e6846133c3c379d3cf94113bd8804893f"},"source":{"id":"2209.05433","kind":"arxiv","version":2},"verdict":{"id":"65622646-54db-4a02-a96e-a76631551664","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T09:42:41.400605Z","strongest_claim":"We demonstrate the efficacy of the FP8 format on a variety of image and language tasks, effectively matching the result quality achieved by 16-bit training sessions. Our study covers the main modern neural network architectures - CNNs, RNNs, and Transformer-based models, leaving all the hyperparameters unchanged from the 16-bit baseline training sessions. Our training experiments include large, up to 175B parameter, language models.","one_line_summary":"FP8 formats E4M3 and E5M2 match 16-bit training accuracy on CNNs, RNNs, and Transformers up to 175B parameters without hyperparameter changes.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen E4M3 and E5M2 encodings will preserve accuracy across all tasks and model scales without any hyperparameter retuning or task-specific adjustments.","pith_extraction_headline":"FP8 with E4M3 and E5M2 encodings matches 16-bit training accuracy on large language and image models without hyperparameter changes."},"references":{"count":26,"sample":[{"doi":"","year":2021,"title":"Michael J. Anderson, Benny Chen, Stephen Chen, Summer Deng, Jordan Fix, Michael Gschwind, Aravind Kalaiah, Changkyu Kim, Jaewon Lee, Jason Liang, Haixin Liu, Yinghai Lu, Jack Montgomery, Arun Moorthy,","work_id":"f1a0c3ad-b21b-473f-a72e-8f0ebd320c2f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"ba7f3302-4a4e-4bc1-85c2-b07ce5a54ccb","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Bfloat16 processing for neural networks","work_id":"b2b017b2-573e-4349-9022-7eb14d1228b7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashc","work_id":"81cf3668-3ff3-4308-a45a-5415987c08ac","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2015,"title":"Binaryconnect: Training deep neural networks with binary weights during propagations","work_id":"07fa38a3-f9ab-416c-bfde-b731f877260c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":26,"snapshot_sha256":"a1347b5bf9598f44b8373eaf98db0aff74d092b150a5f78d5dcdf82406aedd1c","internal_anchors":7},"formal_canon":{"evidence_count":1,"snapshot_sha256":"ce77c6a6004888e65152fb50e6bad0336fd153b773deac150eca9e7e05f3da7d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}