Tesla AI day debut: FSD ultimate evolution? There are also supercomputing Dojo, 50 billion transistor D1 chips and humanoid robots

【TechWeb】 Beijing time. 8 month 20 Japan , Tesla successfully held in the United States AI Day Activities . tesla CEO Elon · musk 、 Director of AI Department Andrej Karpathy And many other engineers , Explain Tesla's pure vision solution online FSD The progress of the 、 Neural network automatic driving training 、D1 chip 、Dojo Supercomputers and other related information .

Implementation of Tesla's pure vision sensor scheme , Multi task learning is indispensable HydraNets Neural network architecture . Every Tesla car has 8 Surround the body 、 Cover all around 360° 's camera , To get traffic lights 、 Signboards 、 ramp 、 Curb and other surrounding information , It provides excellent conditions for neural network learning .

Andrej say ：“ We hope to build a neural network connection similar to the animal visual cortex , Simulate the process of brain information input and output . It's like light entering the retina , We hope to simulate this process through the camera .”

Multi task learning HydraNets Neural network architecture can 8 The images obtained by the two cameras are spliced together , And perfectly balance the delay and accuracy of the video picture . Mark lanes manually or automatically 、 vehicle 、 Signal lamp 、 Obstacles and other environment and dynamic and static objects , The system will analyze the video frame by frame , Understand the depth of an object 、 Speed and other information , Then give these data to the team to learn .

But in the process , Tesla found several problems ： These parameters and spatial tracking are difficult to pass C++ This infrastructure enables splicing ; The output quality of some spatial data is not high ; Different cameras get different object information , It is difficult to grasp the whole when assembling .

To solve these problems , Tesla developed “ Vector space ”（Vector Space） technology , At the same time, it also has a nonconvex optimization algorithm （Non-convex）、 Two advantages of high dimension . The technology can be realized by 8 Draw based on the data input by multiple cameras 3D Aerial view , formation 4D Of space and time tags “ pattern of road distribution ” To present information such as roads , Help the vehicle grasp the driving environment , Find the best driving path more accurately .

With a huge amount of 、 Accurate video data , Tesla also needs to create a powerful neural network , And make a special layout of the network , So that these data can be integrated and re analyzed on a total backbone network . therefore , tesla “ Tall buildings rise from the ground ”, Independently developed a training method based on neural network .

Tesla has a data annotation team composed of talents from all over the world , The scale is 1000 Left and right . The team works on the objects in the video data every day “ Vector space ” Label in , With the cooperation of manual annotation that is good at grasping details and automatic annotation with higher efficiency , Only need to mark once ,“ Vector space ” It can automatically mark multiple frames of all cameras . This brings tens of billions of effective and diversified native data to Tesla , These data will be used for neural network training .

meanwhile , Tesla also developed “ Simulation scene technology ”, It can simulate the less common “ Edge scenes ” For automatic driving training . In the simulation scenario , Tesla engineers can provide different environments and other parameters （ obstacle 、 Collision 、 Comfort, etc ）, It greatly improves the training efficiency .

thus , tesla FSD The system can realize every 1.5 millisecond 2500 Super efficiency of this search , Predict what may happen , And find the safest 、 Most comfortable 、 The fastest autopilot path .

The present , As the data to be processed begins to grow exponentially , Tesla is also improving the computational power of training neural networks , therefore , Tesla Dojo supercomputer .

Tesla's goal is to achieve the ultra-high computing power of artificial intelligence training , At the same time, expand the bandwidth 、 Reduce the delay 、 Cost savings . This requires that Dojo The layout of supercomputers , To achieve the best balance between space and time .

form Dojo The key unit of supercomputer , It is a neural network training chip independently developed by Tesla ——D1 chip .D1 The chip adopts distributed structure and 7 Nanotechnology , carrying 500 100 million transistors 、354 Training nodes , The internal circuit alone is as long as 17.7 km , It realizes super computing power and ultra-high bandwidth .

1500 individual D1 Chip total 53 More than 10000 training nodes , Make up the Dojo Supercomputer training module . Because each D1 The chips are seamlessly connected , The delay between adjacent chips is very low , The training module realizes the bandwidth reservation to the greatest extent , With the high bandwidth created by Tesla 、 Low latency connector , Up to 9PFLOPs（9 Billions of times ）.

Thanks to the independent operation ability and unlimited link ability of the training module , Composed of Dojo The performance expansion of supercomputers is theoretically unlimited , It's a true “ Performance beast ”. Practical application , Tesla will 120 A training module is assembled into ExaPOD, It is the world's leading artificial intelligence training computer . Compared with other products in the industry , Its performance is improved at the same cost 4 times , Performance improvement under the same energy consumption 1.3 times , Space savings 5 times .

Matching powerful hardware , It is a distributed system developed by Tesla ——DPU（Dojo Processing Unit）.DPU Is a visual interactive software , The scale can be adjusted according to requirements at any time , Efficiently process and calculate , Data modeling 、 Storage allocation 、 Optimize the layout 、 Partition expansion and other tasks .

soon , Tesla is about to start Dojo The first assembly of supercomputers , And from the entire supercomputer to the chip 、 System , Further improvement .

In addition to the highly anticipated neural network learning and Dojo supercomputer , At the end of the activity , Musk is talking about AI The direction of development , And out of everyone's expectation “Tesla Bot”.Tesla Bot high 1.72 rice , heavy 56.6 kg , The screen on the face can display information , Have human level hands , And strong feedback , To achieve balanced and agile movements .

Musk said ,Tesla Bot Will make use of Dojo Supercomputer training mechanism to improve function , And added ：“ There will be no shortage of labor in the future , But manual labor is only an option .Tesla Bot Can perform some dangerous 、 Repeatability 、 A boring task .”Tesla Bot Or the first prototype will be launched next year .

Imagine , If half of the human workers on Tesla's production line are replaced by robots ,Model3、Y Will there be further price cuts , Fall into 20 Within ten thousand ？