[{"data":1,"prerenderedAt":143},["ShallowReactive",2],{"project-21877210":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":5,"maker":10,"posted_at":11,"year":12,"points":13,"comments":14,"rank_in_year":15,"shown_in_year":16,"top":17,"tier":18,"category":19,"subcategory":20,"format":21,"keywords":22,"reception":27,"sentiment":28,"stance_mix":29,"scored_comments":31,"aspects":32,"strip":34,"post_text":28,"what":35,"said":69,"maker_in_thread":90,"shape":95,"grid":98,"attempts":105,"same_week":106,"maker_shows":135,"maker_hidden":136,"reader_version":137,"card":138},"21877210","pastebin.com","I trained a GPT-2 1.5B Subreddit Simulator in 3 days using 88 TPUs","title","Show HN: I trained a GPT-2 1.5B Subreddit Simulator in 3 days using 88 TPUs","https:\u002F\u002Fpastebin.com\u002F5qKQs2Hk","sillysaurusx",1577255355,2019,10,3,1798,2195,82,"shown","ai","language model training","other",[23,24,25,26],"gpt-2","tpu training","language model","subreddit simulator","quiet",null,{"positive":30,"neutral":31,"mixed":30,"critical":30},0,1,{"pricing":33},{"positive":30,"critical":30},[30,14,30],{"problem":36,"audience":28,"pricing":43,"motivation":28,"time_to_build":49,"stack":54,"openness":67,"stage":68,"team_size":68},{"value":37,"source":38,"evidence":39},"Training a large GPT-2 model cheaply using many preemptible TPUs instead of a TPU pod","maker",[40],{"id":41,"by":10,"text":42},"21877232","This was done using preemptible TPUv3's, but you can get similar results with TPUv2's. It would just take about 3 times longer. I'm excited about the multi-TPU technique. We've…",{"value":44,"source":38,"evidence":45},"Free to run via TFRC; about $4k per day without TFRC",[46],{"id":47,"by":10,"text":48},"21964159","Nothing, thanks to TFRC. https:\u002F\u002Fwww.tensorflow.org\u002Ftfrc Without TFRC it's about $4k per day. Which is orders of magnitude less expensive than a TPU pod.",{"value":50,"source":38,"evidence":51},"Three days to train",[52],{"id":53},"post",[55,59,63],{"value":56,"source":38,"evidence":57},"GPT-2",[58],{"id":41,"by":10,"text":42},{"value":60,"source":38,"evidence":61},"TensorFlow",[62],{"id":47,"by":10,"text":48},{"value":64,"source":38,"evidence":65},"Python",[66],{"id":41,"by":10,"text":42},"open_source","unknown",{"praise":70,"objections":75,"requests":76,"questions":77,"alternatives":85},[71],{"value":72,"source":38,"evidence":73},"Swarm training across TPUs is cheaper than a TPU pod",[74],{"id":41,"by":10,"text":42},[],[],[78],{"value":79,"source":80,"evidence":81},"How much did the parallel training cost?","commenter",[82],{"id":83,"by":84,"text":79},"21892516","sansnomme",[86],{"value":87,"source":38,"evidence":88},"TPU pod",[89],{"id":41,"by":10,"text":42},{"comments":91,"answered":31,"key":92},2,[93,94],{"id":41,"text":42},{"id":47,"text":48},{"commenters":91,"max_depth":14,"maker_replies":91,"first_comment_mins":96,"lifespan_hours":97},7,279,[99,101,103],[41,30,30,100,10,31],"project\u002Fpositive\u002Fperformance",[83,30,14,102,84,30],"project\u002Fneutral\u002Fpricing",[47,30,30,104,10,31],"project\u002Fpositive\u002Fpricing",[],[107,112,117,121,126,130],{"id":108,"name":109,"points":110,"category":111},"21858226","Angeldust",716,"games",{"id":113,"name":114,"points":115,"category":116},"21855986","22120",406,"infrastructure",{"id":118,"name":119,"points":115,"category":120},"21874772","Gameslab","hardware",{"id":122,"name":123,"points":124,"category":125},"21880446","Cats of Jasnah",367,"learning_science",{"id":127,"name":128,"points":129,"category":125},"21851495","Mandarin Tone Trainer",237,{"id":131,"name":132,"points":133,"category":134},"21841368","gc",192,"languages",5,false,"v2",{"id":139,"name":5,"line":6,"ai":136,"cat":19,"year":12,"points":13,"comments":14,"rec":27,"s":28,"top":17,"kw":140,"maker":10,"domain":5,"tier":18,"attempts":30,"strip":141,"url":9,"mix":142,"praise":72,"objection":-1,"sub":20},21877210,[23,24,25,26],"030",[30,31,30,30],1794036584341]