[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-better-4-bit-recipe-for-training-giant-ai-models":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},1688,"a-better-4-bit-recipe-for-training-giant-ai-models","A Better 4-bit Recipe for Training Giant AI Models","Researchers propose UFP4, a uniform 4-bit training format that avoids a systematic rounding flaw baked into the FP4 standard most GPU makers are shipping.","A new pretraining recipe challenges the 4-bit floating-point format that NVIDIA and AMD have built into their latest AI accelerators.\n\nResearchers publishing on arXiv identified a flaw they call Shrinkage Bias in the E2M1 data format — the 4-bit standard baked into NVIDIA Blackwell and Rubin-class hardware, as well as AMD's MI350 GPUs. The bias stems from geometric asymmetry in how E2M1 represents numbers: rounding errors consistently skew negative, and those errors compound across every layer of a neural network. The paper also shows that the Random Hadamard Transform, a technique commonly used to improve quantization quality, actually amplifies this bias rather than canceling it — explaining instability that practitioners have observed but not fully diagnosed.\n\nThe alternative the researchers propose, UFP4, swaps E2M1 for a uniform 4-bit grid (E1M2 or INT4-style) that sidesteps the geometric problem entirely. Tested on models ranging from a dense 1.5-billion-parameter network to a mixture-of-experts model at 124 billion parameters, UFP4 consistently produced smaller loss degradation relative to full-precision BF16 training than comparable E2M1 baselines. That matters because 4-bit training is one of the few remaining levers for cutting memory and compute costs without shrinking the model itself.\n\nNVIDIA and AMD have already committed silicon to E2M1 as a first-class format — meaning any course correction runs into hardware that is already shipping. The paper's implicit ask — that future accelerators treat uniform 4-bit grids as equals to E2M1 — is a reasonable one, but it arrives after the molds were cast.","[\"ai\",\"machine-learning\",\"hardware\",\"research\"]","2026-06-19T04:00:00.000Z","2026-06-19T09:59:24.828Z","2026-06-19T14:21:37.093Z","published",null,[],"ai",[24,26,27,28],"machine-learning","hardware","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.20381",0,{"sections":35},[36,39,43,48,53,57,62,66,70,75,80,85,90,95],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",490,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",132,{"name":44,"slug":45,"count":46,"latest_published_at":47},"Policy","policy",88,"2026-06-16T09:26:09.000Z",{"name":49,"slug":50,"count":51,"latest_published_at":52},"Consumer Tech","consumer-tech",78,"2026-06-16T17:58:24.000Z",{"name":54,"slug":27,"count":55,"latest_published_at":56},"Hardware",62,"2026-06-18T15:24:16.000Z",{"name":58,"slug":59,"count":60,"latest_published_at":61},"Deals","deals",58,"2026-06-19T14:43:50.000Z",{"name":63,"slug":64,"count":60,"latest_published_at":65},"Software","software","2026-06-16T20:00:00.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":18},"Dev Tools","dev-tools",50,{"name":71,"slug":72,"count":73,"latest_published_at":74},"Science","science",38,"2026-06-18T04:00:00.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Gaming","gaming",31,"2026-06-16T15:25:13.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"General","general",26,"2026-06-13T18:35:15.000Z",{"name":86,"slug":87,"count":88,"latest_published_at":89},"Startups","startups",23,"2026-06-16T15:00:00.000Z",{"name":91,"slug":92,"count":93,"latest_published_at":94},"Reviews","reviews",19,"2026-06-14T08:00:00.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]