Tesla needs massive computing power in many areas, one of the most important of which is artificial intelligence. With its help, Autopilot (Tesla’s autonomous vehicle control system) marks a huge amount of data that the company obtains directly from cars driving in traffic. This is the only way to quickly improve the system.
It will provide the necessary computing power dojo, the successor to the fifth most powerful supercomputer in the world today.
Tesla is building a supercomputer on the Nvidia platform. The main performance consists of the latest generation A100 calculation cards, each with an 80 GB HBM2 memory with a capacity of 2 TB / s.
The current (mass) supercomputer, there are in total 3) consists of 720 knotsEach of them contains eight cards, so in total they make up a supercomputer. 5760 Nvidia A100 computer card with 460 TB of HBM2 memory.
Nvidia A100 computer card
Since a card has a power of 312 TFLOPS (FP16), it represents the total computing power 1.8 EFLOPS. However, it should be noted that this is a lower resolution than the FP16, which is achieved thanks to the tension cores integrated into the chip of the A100 board. However, this precision is sufficient and is used as a standard in artificial intelligence calculations.
High computing power with fast processing also requires fast and large storage. The supercomputer has 10 petabytes Total productivity NVMe SSD 1.6 terabyte / s. The transfer rate between servers is also high 640 TB / s.
Lots of data to process
Tesla representatives revealed that their data set has already grown to 1.5PB and is essentially an image. Elon Musk had previously warned that Tesla would focus on developing autopilot, which would only work with the help of cameras, like a human using eyes. Lidar is a bad word for Elon Musk, at least when it comes to cars, because he has no problem with SpaceX’s space units and there he understands their need for, say, precise docking to the Space Station. international (ISS).
Tesla adapts the artificial intelligence neural network to the architecture of this supercomputer and to possible future versions. In the cars themselves, it already has its own accelerometer chips for artificial intelligence, which act on the basis of data acquired from the aforementioned supercomputer.
Rent a supercomputer like Amazon
Elon Musk too plansThat the unused time of the new supercomputer will be rented through the web interface. Amazon succeeded with a similar model, becoming the first major leader and also the largest cloud service provider. Amazon wanted to effectively take advantage of the increased performance of the data centers and started renting them through the cloud. Today, it is a very important revenue pillar of tens of billions of dollars, and Amazon is the global cloud revenue leader.
Tesla will likely go very similarly and rent unused energy and time, but this system can grow because Amazon was able to create a new type of supercomputer cloud that specializes only in high performance in learning AI. .
If it attracts, for example, other automakers and competing manufacturers of autonomous and other technologies, that could be another income for Tesla. At the same time, it will allow the company to build larger and more powerful supercomputers, which will further lead to the improvement of autonomous vehicle control technology.
The most powerful supercomputer
The Top500.org site already publishes the current ranking The most powerful supercomputer in June 2021, the Japanese rookie Fugaku from Fujitsu, with a power of 537 PFLOPS for a consumption of almost 30 megawatts, remains in first place. In second position comes the well-known American top IBM with a power of 200 PFLOPS with a consumption of 10 MW and the three are also owned by the American supercomputer Sierra from IBM with a maximum power of 126 PFLOPS and a consumption of 7 , 6 MW. .
An updated overview of the world’s most powerful supercomputers
In fourth place comes the Chinese Sunway TaihuLight (90 PFLOPS, 15.3 MW), and in fifth place the American Perimutter with AMD Epyc 7763 processors and Nvidia A100 computer cards (older with 40 GB of memory).
As we said before, there are plenty of supercomputers in development that should surpass the magic limit of EXAFLOPS, in this case high-resolution computing power for the most precise scientific calculations.
Tesla Autopilot Basics