**有了自己的“CUDA”工具箱

口罩哥KZG

10-01

DeepSeek开源六大组件，对标英伟达

2026.10.01

一一对应四个字，是这件事的全部分量。

9月30日，DeepSeek官方公众号发文。面向****，开源全套基础设施组件….

关键是对应关系。

这套组件与此前英伟达平台那套，一一对应。TileLang封装Ascend C指令。性能不打折。

六件套分工明确。DeepGEMM算矩阵，DeepEP管通信。TileKernels做访存。FlashMLA做稀疏注意力。DeepSelect筛数据。

多项测试里，计算通信性能接近硬件上限。每个TileLang算子，**上都有高性能实现。

硬件侧，128颗**950超节点推进。但950DT年产几十万颗。16万颗大单交付排一年以上。

黄仁勋4月讲过。

DeepSeek先把**芯片跑起来。对美国是糟糕的结果。节后盯交付采用率。要开始围殴cuda了…

免责声明：上述内容仅代表发帖人个人观点，不构成本平台的任何投资建议。

精彩评论

我们需要你的真知灼见来填补这片空白

发表看法

{"i18n":{"language":"zh_CN"},"detailType":1,"isChannel":false,"data":{"magic":2,"id":613702607605872,"tweetId":"613702607605872","gmtCreate":1790849400000,"gmtModify":1790852242814,"author":{"id":172401613997661,"idStr":"172401613997661","authorId":172401613997661,"authorIdStr":"172401613997661","name":"口罩哥KZG","avatar":"https://static.tigerbbs.com/c75702f7262e33dac5081a8655b23d87","vip":6,"userType":6,"introduction":"抖音财经百万大V “左达右饼”创始人","boolIsFan":false,"boolIsHead":false,"crmLevel":7,"crmLevelSwitch":1,"individualDisplayBadges":[],"wearingBadges":[],"fanSize":394,"starInvestorFlag":false},"themes":[],"images":[{"img":"https://static.tigerbbs.com/8119030b6b164475bb18dc150f8655ac"}],"coverImages":[{"img":"https://static.tigerbbs.com/8119030b6b164475bb18dc150f8655ac"}],"title":"**有了自己的“CUDA”工具箱","html":"<html><head></head><body><div>\n DeepSeek开源六大组件，对标英伟达\n 2026.10.01\n <blockquote>\n 一一对应四个字，是这件事的全部分量。\n </blockquote>\n 9月30日，DeepSeek官方公众号发文。面向****，开源全套基础设施组件….\n 关键是对应关系。\n 这套组件与此前英伟达平台那套，一一对应。TileLang封装Ascend C指令。性能不打折。\n 六件套分工明确。DeepGEMM算矩阵，DeepEP管通信。TileKernels做访存。FlashMLA做稀疏注意力。DeepSelect筛数据。\n 多项测试里，计算通信性能接近硬件上限。每个TileLang算子，**上都有高性能实现。\n <img src=\"https://static.tigerbbs.com/8119030b6b164475bb18dc150f8655ac\">\n 硬件侧，128颗**950超节点推进。但950DT年产几十万颗。16万颗大单交付排一年以上。\n 黄仁勋4月讲过。\n DeepSeek先把**芯片跑起来。对美国是糟糕的结果。节后盯交付采用率。要开始围殴cuda了…\n</div></body></html>","htmlText":"<html><head></head><body><div>\n DeepSeek开源六大组件，对标英伟达\n 2026.10.01\n <blockquote>\n 一一对应四个字，是这件事的全部分量。\n </blockquote>\n 9月30日，DeepSeek官方公众号发文。面向****，开源全套基础设施组件….\n 关键是对应关系。\n 这套组件与此前英伟达平台那套，一一对应。TileLang封装Ascend C指令。性能不打折。\n 六件套分工明确。DeepGEMM算矩阵，DeepEP管通信。TileKernels做访存。FlashMLA做稀疏注意力。DeepSelect筛数据。\n 多项测试里，计算通信性能接近硬件上限。每个TileLang算子，**上都有高性能实现。\n <img src=\"https://static.tigerbbs.com/8119030b6b164475bb18dc150f8655ac\">\n 硬件侧，128颗**950超节点推进。但950DT年产几十万颗。16万颗大单交付排一年以上。\n 黄仁勋4月讲过。\n DeepSeek先把**芯片跑起来。对美国是糟糕的结果。节后盯交付采用率。要开始围殴cuda了…\n</div></body></html>","text":"DeepSeek开源六大组件，对标英伟达 2026.10.01 一一对应四个字，是这件事的全部分量。 9月30日，DeepSeek官方公众号发文。面向****，开源全套基础设施组件…. 关键是对应关系。 这套组件与此前英伟达平台那套，一一对应。TileLang封装Ascend C指令。性能不打折。 六件套分工明确。DeepGEMM算矩阵，DeepEP管通信。TileKernels做访存。FlashMLA做稀疏注意力。DeepSelect筛数据。 多项测试里，计算通信性能接近硬件上限。每个TileLang算子，**上都有高性能实现。 硬件侧，128颗**950超节点推进。但950DT年产几十万颗。16万颗大单交付排一年以上。 黄仁勋4月讲过。 DeepSeek先把**芯片跑起来。对美国是糟糕的结果。节后盯交付采用率。要开始围殴cuda了…","highlighted":1,"essential":1,"paper":2,"likeSize":0,"commentSize":0,"repostSize":0,"favoriteSize":0,"link":"https://laohu8.com/post/613702607605872","repostId":0,"isVote":1,"tweetType":1,"viewCount":1859,"commentLimit":10,"likeStatus":false,"favoriteStatus":false,"reportStatus":false,"symbols":["CLSE","IETC","FT_SY_SNVDA","3NVD.UK","SPTE","ARTY","FAI","07788","07388","TEK","NVD","IXN","NVII","LFGY","FTEC","CMAG","ILDR","NULC","TTXU","NVDA","NVDB","NVDD","NVDL","NVDG","ROM","NVDQ","NVD2.UK","SOXQ","09788","NVDS","CHPX","09388","ABFL","NVDY","NVDU","SOXX","NVDW","NVDX","SNVD.UK","NVDS.UK","ANV","CHAT","CNEQ","WISE","TTEQ","NVD3.UK","2NVD.UK","XLK","JGRO","NVIW.SI"],"verified":2,"subType":0,"readableState":1,"langContent":"CN","currentLanguage":"CN","warmUpFlag":false,"orderFlag":false,"shareable":true,"causeOfNotShareable":"","featuresForAnalytics":[],"commentAndTweetFlag":false,"andRepostAutoSelectedFlag":false,"upFlag":false,"length":561,"optionInvolvedFlag":false,"subscribersOnly":false,"subscribersOnlyAccessible":false,"xxTargetLangEnum":"ZH_CN"},"commentList":[],"isCommentEnd":true,"isTiger":false,"isWeiXinMini":false,"url":"/m/post/613702607605872"}