特斯拉AI日：发布自研超算Dojo芯片，7nm制造，每秒可执行1024亿次计算

{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"D1 芯片来了，Dojo 还会远吗？"}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"特斯拉发布 Dojo D1 芯片"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"北美时间 8 月 19 日，特斯拉 AI Day 正式举行。在发布会上，马斯克介绍了特斯拉在人工智能领域的软件和硬件进展。其中，自研超级计算机 "},{"type":"link","attrs":{"href":"https:\/\/www.cnbc.com\/2021\/08\/19\/tesla-unveils-dojo-d1-chip-at-ai-day.html","title":"xxx","type":null},"content":[{"type":"text","text":"Dojo D1 芯片"}]},{"type":"text","text":"成为本次发布会最大的亮点。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/05\/b1\/05e7570228f7db6599c5b5ecedc27db1.png","alt":null,"title":"Ganesh Venkataramanan 展示 D1 训练芯片","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"会上，Dojo项目负责人Ganesh Venkataramanan上台就Dojo的主要性能进行了展示。Ganesh Venkataramanan表示，马斯克想要一台超快的训练计算机来训练 Autopilot。因此，Project Dojo 诞生了。Dojo架构拥有一个大规模计算平面，极高宽带和低延迟。作为Dojo架构的重要组成部分，D1芯片采用7纳米制造工艺，处理能力为每秒1024亿次。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/3b\/93\/3b9db9817d8f81e9d6225c44ffce6493.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"曾在芯片制造商AMD工作的Venkataramanan表示，将一组这样的芯片放置在单个“训练片”上，以提供每秒九千万亿次的计算能力，并将120个芯片放在多个服务器机柜上，达到每秒超过1千万亿次的计算能力。这些芯片可以帮助训练模型来识别特拉斯汽车摄像头中收集到的各种物品。训练模型需要大量的计算工作。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/58\/ab\/58fd17c8f73be111b91573c85914fbab.jpeg","alt":null,"title":"一个特斯拉 Dojo 机器学习计算块","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Dojo的规格、性能十分强大。Dojo是一个真正意义上的机器学习架构，它拥有50多万个训练节点。每秒九千万亿次的计算以及每秒36tb的带宽还都只是Dojo的冰山一角。比这更可怕的是，Dojo的全部潜力都在为一件事情做准备——使自动驾驶汽车成为可能。Venkataramanan认为，特斯拉的技术将打造出最快的AI训练计算机。他还表示：“特斯拉很快就会组装首批套件。\""}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"此外，马斯克还表示可能会在明年推出名为“特斯拉机器人”的人形机器人原型，这种具有人类外观的机器人将执行人们最不喜欢做的工作，“消除危险、重复、无聊的任务”，并“对经济产生深远的影响”。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/cf\/30\/cf67e9dbfa7dcbc1fa5acb9d4533c330.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"超级计算机 Dojo"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"事实上，大家对于特斯拉本次发布 Dojo 芯片并不意外。在特斯拉本月初发布的邀请函上，一颗大规模芯片模组的结构图格外吸睛，彼时就有不少人猜测，这是特斯拉为超级计算机 Dojo 设计的专用芯片。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/fe\/1d\/fe4c823fd009067743dc162716c9e91d.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"而在特斯拉发布邀请函前两日，自动驾驶领域专家 Dennis Hong 就已在其 Twitter 上发布过这张图片，并打上了 Tesla AI Day 标签。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/7f\/a0\/7f98eff476b1379776eff253190d92a0.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"据了解，Dennis Hong 是加州大学洛杉矶分校（UCLA）机械和航天工程学院的教授、机器人实验室 Romela 负责人，主要研究方向为类人机器人、机器设计以及自动驾驶汽车研究。2007 年，Dennis Hong 及团队参加了美国 DARPA 无人驾驶汽车挑战赛，并获得了第三名的好成绩。他们基于一辆福特 SUV 制造了一款自动驾驶汽车，无需干预即可自动到达目的地，如果按照美国汽车工程学会（SAE）的标准，这辆车的自动驾驶能力可以达到 L4 级。2011 年，Dennis Hong 发表了一个以“为盲人制造汽车”为主题的 TED 演讲，在自动驾驶领域内颇具影响力。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"外媒 "},{"type":"link","attrs":{"href":"https:\/\/www.teslarati.com\/tesla-dojo-chip-images-dennis-hong\/","title":"xxx","type":null},"content":[{"type":"text","text":"Teslarati 推测"}]},{"type":"text","text":"，Dennis Hong 可能已经在私下里与特斯拉成立了合作关系，前者将作为编外专家帮助后者进行 Dojo 超级计算机的调试与完善，不过特斯拉官方和 Dennis Hong 本人都未对此观点表示承认或否认。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"事实上，早在 2019 年的 Autonomy Day 上，马斯克就提到过 "},{"type":"link","attrs":{"href":"https:\/\/www.theverge.com\/2021\/8\/18\/22627101\/tesla-event-ai-day-rumors-elon-musk-what-to-expect","title":"xxx","type":null},"content":[{"type":"text","text":"Dojo"}]},{"type":"text","text":"，并表示 Dojo 能够利用海量的视频（级别）数据，做“无人监管”的标注和训练的超级计算机。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/b7\/5f\/b750aa14e86b0081a08261a0ca50aa5f.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"2020 年 8 月 15 日，马斯克发推称“特斯拉正在开发一款名为 Dojo 的神经网络训练计算机，用于处理大量视频数据”。当时马斯克还暗示了 Dojo 的计算能力，声称它具备 exaFLOP 性能，即每秒能够执行 1018 万亿次的浮点运算。超越日本富岳，成为世界第一。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/resource\/image\/37\/d0\/372751f8a7789f000524990805cf48d0.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"至于为什么要自研超算 Dojo，马斯克也给过答案：“解决自动驾驶的唯一方法是解决现实世界中的 AI 问题，无论是硬件还是软件，而这也是特斯拉正在做的事情。除非一家公司具有很强的 AI 能力以及超强算力，否则很难解决自动驾驶难题。”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如今 Dojo 芯片已来，根据马斯克的计划，明年 Dojo 会投入运营。至于 Dojo 未来会为特斯拉以及自动驾驶带来哪些改变，一起拭目以待。"}]}]}