[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-used-optane-memory-runs-a-28t-parameter-ai-model-on-a-cpu":10,"sections":35},{"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":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},12571,"used-optane-memory-runs-a-28t-parameter-ai-model-on-a-cpu","Used Optane Memory Runs a 2.8T-Parameter AI Model on a CPU","A homelabber got Kimi K3 running on discontinued Optane and a Xeon at under one token per second, and the memory sticks need active cooling.","A homelab builder has run Moonshot AI's 2.8-trillion-parameter Kimi K3 model on a CPU alone, using discontinued second-gen Optane memory, at 0.85 tokens per second.\n\nThe build pairs a 40-core 3rd-gen Xeon Scalable processor with Optane PMem 200 modules, Intel's persistent memory that sits in ordinary RAM slots. That speed comes with 1-bit quantization, a compression trick that stores a model's numbers in fewer bits. At 3-bit, it fell to 0.75 tokens per second. The builder used App Direct Mode, which exposes Optane as a separate region the software maps directly, instead of Memory Mode, where DRAM acts as a cache in front of Optane.\n\nAn earlier build by a different user reached roughly 4 tokens per second, but the comparison is loose. That machine used first-gen Optane in Memory Mode, an RTX 3060 graphics card, 2-bit quantization and the smaller 1-trillion-parameter K2.5 model.\n\nHeat is the catch. Intel rates each module at 12W or 15W and specifies forced airflow of 2.4 to 2.7 meters per second, a data center condition. Four sticks could put up to 60W next to the system memory. Optane throttles at 83 degrees Celsius and trips a shutdown at 85, while typical DDR4 server memory is rated to 85 and up to 95 in an extended range. The builder aimed dual 12W blower fans at the modules and warned that the heat can endanger neighboring DRAM.\n\nThe appeal is capacity, not speed. Moonshot recommends data center inference engines for Kimi K3, and even the smallest community quantization needs at least 610GB of RAM. With DRAM prices up sharply, the builder says used Optane costs a fraction of the DRAM equivalent. One commenter estimated a 50k-token task would take 13 hours at about 1 token per second. Another noted that overnight batch jobs do not care. That makes this a tool for patient, offline work rather than chat.\n\nThe numbers are also pre-tuning, with no adjustment of thread counts or NUMA affinity (how work is matched to memory banks). The builder plans to add GPUs and a DRAM tier alongside the Optane. Whether that closes the gap to the first build's 4 tokens per second on a model nearly three times larger is the open question.","[\"optane\",\"local-llm\",\"homelab\",\"memory\"]","2026-11-08T12:20:00.000Z","2026-11-08T13:25:12.227Z","2026-11-08T13:25:12.473Z","published",null,[],"hardware",[26,27,28,29],"optane","local-llm","homelab","memory",[31],{"name":32,"url":33},"Tom's Hardware","https:\u002F\u002Fwww.tomshardware.com\u002Ftech-industry\u002Fartificial-intelligence\u002Fhomelabber-uses-second-gen-optane-memory-sticks-to-run-a-2-8-trillion-parameter-ai-model-on-cpu-alone-warns-that-optanes-extreme-heat-can-burn-out-nearby-memory-dram-used-blower-fans-to-prevent-melting-ram",0,{"sections":36},[37,42,47,52,57,62,66,71,76,81,86,91,96,101],{"name":38,"slug":39,"count":40,"latest_published_at":41},"AI","ai",7110,"2026-11-08T14:34:09.000Z",{"name":43,"slug":44,"count":45,"latest_published_at":46},"Security","security",1099,"2026-11-08T15:24:01.000Z",{"name":48,"slug":49,"count":50,"latest_published_at":51},"Deals","deals",620,"2026-11-08T08:00:00.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":56},"Policy","policy",604,"2026-11-07T15:48:26.000Z",{"name":58,"slug":59,"count":60,"latest_published_at":61},"Consumer Tech","consumer-tech",336,"2026-11-08T21:00:00.000Z",{"name":63,"slug":24,"count":64,"latest_published_at":65},"Hardware",316,"2026-11-08T14:36:39.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Science","science",222,"2026-11-07T16:19:17.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",154,"2026-11-08T15:42:52.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",150,"2026-11-06T16:50:01.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Dev Tools","dev-tools",125,"2026-11-06T22:15:59.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",111,"2026-11-08T21:27:29.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Gaming","gaming",94,"2026-11-08T20:52:05.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",38,"2026-10-20T12:00:00.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]