Development of proprietary chips based onartificial intelligence is a rather complicated process. But Tesla is one of the most progressive companies today and one of the pioneers in creating self-driving cars. Therefore, the company management believes that the forces and means that are invested in such technologies are worth it. And recently, Tesla introduced its new AI chip, which is 21 times faster than the Nvidia model used in Tesla cars now. Well, in order to completely “finish off” everyone, company representatives said that in each new car there would be not one such chip, but two at once.
How is the new chip from Tesla
Tesla CEO Ilon Musk and hiscolleagues showed new equipment (and this, by the way, is the third generation of chips for self-driving cars) back in April. But at a recent Hot Chips conference on Tuesday, developers revealed a ton of detail. In addition to the already mentioned performance increase of 21 times compared to chips from Nvidia, Tesla says that the cost of production is lower than the chips of the previous generation.
The company needed betterequipment to achieve the goal of creating fully autonomous cars, ”said Ganesh Venkataramanan, one of the developers of chips for Tesla and a former AMD processor engineer. It was clear to us that in order to achieve this result, we needed to go beyond the boundaries of the form factor used and develop something of our own.
The number of transistors used whenThe creation of the chip is actually impressive: 6 billion. But the point here, of course, is not in the number of transistors, but in architecture. And in the production facilities of the company. Each Tesla trip computer will have two AI chips. The second is installed to provide greater security. In addition, each chip will have its own power supply in order to protect the chips from failures.
It is interesting: The safest electric car is named. How do you like that, Elon Musk?
Each chip makes its own assessment ofwhat the car should do next. The computer compares the data, and if they match, the car does what is planned. If the chips do not agree, the machine tries to analyze the situation again. More recently, we wrote about a similar incident. And if you are subscribed to our news telegram channel, then you are probably in the know about this event.
Each Tesla chip is clocked at 2 GHz andperforms 36 trillion operations per second. This performance is possible because Tesla optimized chips for self-driving cars and directed all the power to processing data that is important for autonomous control. High performance was achieved by optimizing floating point operations by writing data not in 16-bit, but in 8-bit format. Such data is read faster and, accordingly, written faster.
The chip also has 32 megabytes of high speedSRAM, which means that it doesn’t have to wait while retrieving data from the much slower conventional DRAM that is used in similar chips today.
Instead of wasting all the “chip power" ondecryption of data and its extraction, we wanted to direct productivity to really important processes, ”says Das Sarma, a member of the new chip development team.
To develop a chip based on artificialintelligence took 14 months, and now everything is ready to arrange its production. Samsung will be responsible for it, the factories of which will be assembled. All new Tesla car models will have at their disposal a high-performance chip. On the old versions of cars assembled at Tesla factories, instead of the previous model, a new device will also be installed. And in the end, all the machines coming off the Tesla conveyor will be running a new chip.