TSMC said its customers are looking for high performance computing to support autonomous driving in vehicles. That's new. And also old.
Among the new process nodes introduced by Taiwan Semiconductor Manufacturing Co. at its 2021 Virtual Technology Symposium was N5A, which the company said is “aimed at satisfying the growing demand for computing power in newer and more intensive automotive applications such as AI-enabled driver assistance and the digitization of vehicle cockpits.”
In a blog post, TSMC head of global marketing Godfrey Cheng, wrote that, “Compared to TSMC’s N7 technology with the Automotive Service Package, N5A delivers a ~20% improvement in performance or a ~40% improvement in power efficiency and a ~80% improvement in logic density.”
What most people skimmed over about those first few amazing autonomous vehicles (AV) is that they all had the rough equivalent of a small data center in the trunk. Ever since, auto companies and their suppliers have been trying to figure out a practical amount of computing power sufficient to support autonomous driving — and that amount has been significantly less than a small data center. But what if, 10 years after those first AVs were demonstrated, putting a supercomputer in a car has actually become practical — or is close enough to consider it?
TSMC senior vice president of business development Kevin Zhang carved out a few minutes from his Technology Symposium schedule for a one-on-one interview with EE Times. In our brief conversation, we asked Zhang about TSMC’s automotive business. He said TSMC is seeing demand for HPC (high performance computing) from automotive makers. They want the performance, power efficiency, and logic density associated with HPC that can be achieved at smaller processing nodes.
That performance must be balanced, the company noted, against the “stringent quality and reliability requirements of AEC-Q100 Grade2as well as other automotive safety and quality standards.”
The automotive semiconductor products that currently satisfy those requirements tend to map to so-called legacy nodes — 40nm, 28nm, etc., far from the TSMC’s stomping grounds at the leading edge — 7nm, 5nm, and soon 3nm.
That’s part of the reason why the automotive market represents only about 4 percent of TSMC’s overall business, according to Zhang. TSMC is currently making mostly microcontrollers with embedded memory for automotive, he said.
But now some automotive makers are looking for HPC in automotive to support autonomous driving, Zhang said. You can do a lot in the cloud, he explained, but if a vehicle has to make a split-second decision, it needs high performance, highly efficient processing right there at the edge.
“This is going to be an important part of future autonomous capability,” Zhang added.
While there seems to be growing demand for far more intelligence at the edge (in this case, in the vehicles themselves), there seems to be no agreement among TSMC’s customers what form it will take. Zhang said different customers have different approaches for supporting AI at the edge; some are using traditional processors, some want ASICs, and some have other approaches.
Whichever way, it seems that AVs are going to be another case of something old becoming new again.