[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-deepx-claims-100b-parameter-inference-under-5-w-on-new-chip":10,"sections":41},{"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":30,"tags":31,"sources":36,"feedback":40,"feedback_at":22,"cost_usd":40,"total_tokens":40},12651,"deepx-claims-100b-parameter-inference-under-5-w-on-new-chip","DeepX Claims 100B-Parameter Inference Under 5 W on New Chip","DeepX says its second-generation chip handles models up to 100B parameters under 5 W using in-memory processing, with a projected 6.7x gain.","DeepX says its second-generation AI chip can run inference on models of up to 100B parameters while drawing less than 5 W. The company has taped out the design, meaning it is finished and sent off for manufacturing. It projects a 6.7x performance gain from pairing the chip with LPDDR5X memory that does some of the computing itself, a technique called processing-in-memory, or PIM.\n\nTwo parts of that claim are softer than they sound. Tape-out is not working silicon in customers' hands, and \"projects\" marks the 6.7x as a forecast rather than a measurement. Both the power figure and the performance gain come from DeepX.\n\nThe PIM angle is the interesting part. Running a large language model is mostly a memory problem. To generate each token, a chip has to read essentially all of the model's weights. For scale, 100B parameters at 4-bit precision is roughly 50 GB of weights, and hauling that much data between memory and processor is where much of the energy goes. Doing some of the arithmetic inside the memory cuts that traffic. Memory makers have pitched PIM for years, mostly for data-center hardware. LPDDR5X is a low-power memory type common in mobile devices, so a design built around it is aimed at a different corner of the market.\n\nThe open questions are the ones that decide whether the number is useful. A chip that holds a 100B-parameter model in under 5 W is only practical if it generates text at a tolerable speed, and 6.7x over which baseline is the first thing to pin down. Until DeepX answers both, the claim describes capacity and power, not how the chip performs in use.","[\"ai-chips\",\"edge-ai\",\"inference\",\"processing-in-memory\"]","2026-11-09T18:33:58.000Z","2026-11-09T19:24:10.099Z","2026-11-09T19:24:10.321Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Remove the meta sourcing phrase 'the source gives no baseline for the 6.7x' (say instead that DeepX has not published a baseline, only if that can be verified), and drop the unsupported 'phone-sized power budget' framing in the dek, since the source only supports '\u003C5 W' and 'up to 100B parameters'.","resolved","hardware",[32,33,34,35],"ai-chips","edge-ai","inference","processing-in-memory",[37],{"name":38,"url":39},"Hacker News","https:\u002F\u002Fwww.eetimes.com\u002Fdeepx-tapes-out-second-gen-chip-projects-6-7x-performance-gain-with-lpddr5x-pim\u002F",0,{"sections":42},[43,48,53,58,63,68,72,77,82,87,92,97,102,107],{"name":44,"slug":45,"count":46,"latest_published_at":47},"AI","ai",7144,"2026-11-12T13:27:45.000Z",{"name":49,"slug":50,"count":51,"latest_published_at":52},"Security","security",1128,"2026-11-11T18:05:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Deals","deals",642,"2026-11-12T14:00:00.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Policy","policy",619,"2026-11-12T12:53:36.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",364,"2026-11-12T11:12:32.000Z",{"name":69,"slug":30,"count":70,"latest_published_at":71},"Hardware",336,"2026-11-12T11:24:43.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Science","science",230,"2026-11-12T10:00:00.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Software","software",168,"2026-11-11T17:36:49.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"Startups","startups",159,"2026-11-11T12:30:00.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Dev Tools","dev-tools",127,"2026-11-11T19:00:44.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"General","general",118,"2026-11-11T16:49:42.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"Gaming","gaming",107,"2026-11-12T10:09:03.000Z",{"name":103,"slug":104,"count":105,"latest_published_at":106},"Reviews","reviews",38,"2026-10-20T12:00:00.000Z",{"name":108,"slug":109,"count":110,"latest_published_at":111},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]