Get email delivery of the Cadence blog featured here
Tom Quan recently came to Cadence to talk about TSMC's automotive strategy. Tom and I go back a long way since fifteen years ago, when I was running the custom IC group. I hired him into Cadence to work for me to run a project that had the nickname Superchip, but I won't rehash old history here. Especially since Superchip was anything but super.
Last year in TSMC Technology Update, I talked about how TSMC have moved away from a "mobile only" strategy to one that fully committed to four markets: mobile (still), automotive, HPC, and IoT. By fully committed they meant more than just marketing, but also special IP libraries, special flows, and special process options. Tom came to tell us what was going on in a bit more detail.
The reason for focusing on these three new markets is that mobile is growing only slowly. It is still huge, of course, but automotive and IoT together, what TSMC calls emerging applications, are forecast to grow at 28% CAGR between now and 2020, whereas mobile is only forcecast to grow at 8%. Of course, emerging applications are starting from a smaller base.
Automotive is actually two rather different markets. Traditional automotive is analog heavy, built in ten-year-old processes that have been characterized to death, and is growing relatively slowly. The newer market is focused on advanced driver assistance systems (ADAS) such as lane following, and on infotainment, which covers everything from music to GPS-based maps. ADAS, of course, is a stepping stone on the route to fully autonomous vehicles, which different sources predict will arrive at different times, with most estimates in the early 2020s.
The change in requirements for autonomous driving are quite dramatic. Just between 2015 and this year, the requirements are for twenty times the compute power and a lot more sensors and actuators. See the table above.
Autonomous vehicles are making very fast advances. It is hard to believe, but it is only 12 years ago that in the DARPA challenge, the furthest any vehicle managed was eight miles. Now Google's cars have surpassed 2 million miles (and that was in October so maybe it is nearer 3 million now), Delphi drove a car across the US (although they were only autonomous on freeways and highways) and, of course, Tesla's autopilot is racking up a lot of miles on a daily basis.
Tom emphasized that automotive is not new to TSMC. The above picture shows where some of TSMC's IP ends up in vehicles. They have over 40 customers, over 500 tapeouts and have shipped over 1 million 12" equivalent wafers into automotive markets. On their leading-edge processes, the two that TSMC are targeting for automotive are 16FFC and 7FF. Automotive platforms are already available for 16FFC and are being prepared for 7FF. There are processes in between, but these will not have all the specialized automotive IP and qualification available. One of these is TSMC9000-A for automotive, built on the existing TSMC9000 IP qualification program. This is in place for 16FFC and will presumably be extended to 7FF in the near future.
As part of its automotive enablement, TSMC provides:
The table below shows the portfolio of automotive-grade IP that is available. Blue dots mean that silicon reports are ready, green dots indicate that design kits are ready and the silicon reports will be available in the next quarter or so.
In this post I've focused solely on TSMC's specialized automotive offering. Of course, lots of their regular technology such as InFO, CoWoS, and the 7nm production ramp schedule, also applies to automotive. Tom covered some of this but this post is long enough already.
Not directly to do with TSMC (although I would be surprised if some TSMC silicon is not in there somewhere), the most impressive video I have seen of autonomous driving is this one in San Francisco (which actually drives past where I live in the middle). It is GM and GM's recently acquired Cruise Automation. Assuming they are being honest that this is fully autonomous, it is much more impressive than other videos I've seen, such as Tesla driving around the Stanford campus. The video shows 13 minutes of driving but it is sped up to five times normal speed so it is just two and a half minutes long. It is worth a watch to see where the state-of-the-art is today.