According to a report by Taiwan media Chinatimes, TSMC will manufacture the latest chip developed by IC design leader Broadcom and Tesla, with several important details:
- The chip is jointly developed by Broadcom and Tesla.
- The chip will be manufactured using TSMC’s 7 nm process.
- The chip uses SoW technology for packaging, which can directly integrate HPC chips with heat dissipation modules without the need for substrates and PCBs.
- The chip is expected to start production in Q4 2020, with an initial production scale of 2,000 pieces.
According to Gartner, compared with 16 nm/14 nm chips, 7 nm chips have increased speed by 35%, reduced power consumption by 65%, and increased density by 3.3 times.
In April 2019, Tesla released its HW 3.0 hardware, which uses a 14 nm process, with a total computing power of 144 Tops and power consumption of 72 W.
This 7 nm process chip will most likely become the core hardware of Tesla’s next-generation Autopilot, which is HW 4.0.
Looking at the entire industry, Mobileye EyeQ5 and Nvidia’s Drive AGX Orin, both released in May 2020, use a 7 nm process.
Obviously, the 14 nm process can no longer meet the high demands for computing power and power consumption in automatic driving. Prior to this, car-grade chips often used more mature technology to meet more demanding working conditions to ensure absolute safety. Therefore, car-grade chips are usually several generations behind consumer-grade chips. However, with the acceleration of intelligence, this phenomenon has slowly begun to change.
On August 15, Elon revealed on Twitter that Tesla is developing Dojo for self-supervised learning to train neural networks. At the same time, Tesla’s deep neural networks are moving from 2D to 4D (adding a time dimension on the basis of 3D), so the demand for computing power will also increase exponentially. So will hardware reach the bottleneck first or will software be unable to fully utilize the hardware?
Regardless of the outcome, such reckless advancement will definitely benefit consumers.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.