To develop its own chips based on artificial intelligence is a pretty complex process. But Tesla is one of the most progressive companies today and one of the pioneers in the field of creating self-driving cars. The company therefore believes that power and money is invested in similar technologies, are worth it. And Tesla recently unveiled its new AI chip, which is 21 times faster than the model from Nvidia, used in cars Tesla now. But to finally all to “finish” the company representatives said that in every new car so the chip will be not one, but two.
How does the new chip from Tesla
Executive Director of Tesla Ilon Musk and his colleagues showed the new equipment (which, by the way, is the third generation of chips for self-driving cars) in April. But at the recent Hot Chips conference on Tuesday, the developers have told a lot of details. Apart from the already mentioned productivity in 21 times in comparison with chips from Nvidia, Tesla claim that the cost of production is lower than that of the chips of the previous generation.
The company needed better equipment to achieve completely Autonomous vehicles, said Ganesh Venkataramanan, one of the developers of chips for Tesla and a former engineer AMD. It was clear that in order to achieve this result, we need to go beyond the boundaries of the used form-factor, and develop something of their own.
The number of transistors used in the creation of the chip, really impressive: 6 billion. But the point here is, clearly, not the number of transistors and architecture. And production capacity of the company. Each onboard computer Tesla will have two chip AI. The second is set to provide greater security. Each chip will have its own power source in order to protect the chip from glitches.
Each chip does its own assessment of what 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 re-analyze the situation. Recently we wrote about a similar incident.
Each Tesla chip operates at a frequency of 2 GHz and perform 36 trillion operations per second. This performance is possible because Tesla optimized chips for self-driving cars and sent all the power to the data processing, which is important for Autonomous control. High performance was achieved through optimization of floating-point operations by writing data in 16-bit and 8-bit format. These data are read faster and, therefore, quickly written.
Also the chip has 32 megabytes of high-speed SRAM, which means it does not need to wait during the retrieval of data from a much slower conventional memory DRAM that is used in such chips today.
Instead of spending all the “power chip” for decryption to retrieve them, we wanted to send the performance on the really important processes, says Das Sarma, member of the team developing a new chip.