• Skip to main content
  • Skip to search
  • Skip to footer
Cadence Home
  • This search text may be transcribed, used, stored, or accessed by our third-party service providers per our Cookie Policy and Privacy Policy.

  1. Blogs
  2. Breakfast Bytes
  3. EDPS 2019: Efficient Design and Manufacturing
Paul McLellan
Paul McLellan

Community Member

Blog Activity
Options
  • Subscribe by email
  • More
  • Cancel
electronic design process symposium
EDPS
ARM

EDPS 2019: Efficient Design and Manufacturing

21 Oct 2019 • 5 minute read

 breakfast bytes logoEDPS, the Electronic Design Process Symposium, took place on October 3 and 4 at SEMI HQ in Milpitas. The theme was Efficient Design and Manufacturing.

Camille Kokozaki

The meeting opened with a moment of silence for Camille Kokozaki, who had helped organize EDPS over the last few years and who died earlier in the year while on a business trip to France. Herb Reiter talked about meeting him when he was running VLSI Technology's design center in Florida. But I met him a few years earlier. VLSI signed an agreement with Olivetti to get access to our technology and that involved putting a team of three employees onsite in Ivrea, where Olivetti was based (it's about an hour north of Turin). They were Yves Saboret, Albert Stritter...and Camille. For reasons I don't remember anymore, I visited for a couple of days in about 1985, a year or two before I would relocate to France and set up VLSI Technology's European R&D site in Sophia Antipolis. Yves, Albert, and Camille all remained at VLSI for well over a decade afterward, as did I, so I continued to run across him from time to time, especially once he moved out to Silicon Valley. When Compass was sold to Avant! and VLSI to Philips/NXP, we all moved on. There is something especially sad about dying alone in a hotel room on another continent from your family. The first time I met his family, as it happens, was the memorial service held for him this summer. Rest in peace, Camille.

Arm's Rob Aitken

The opening keynote was by Rob Aitken of Arm. He titled his presentation How Different Is ADAS? ADAS, just in case you've been on a few years' sabbatical off the grid, stands for advanced driver assistance systems. This term covers everything from automatic emergency braking (AEB, mandated for all vehicles in the US by 2022 and in Europe by next year) to adaptive cruise control and lane following, up to higher levels of automation where it starts to be called autonomous driving. For a quick tutorial on all things automotive, see my post Automotive Industry Basics.

Rob started with the dream:

But brought everyone down to earth by pointing out that this wasn't going to happen. He pointed out that speed is expensive: more energy, more complexity, less margin (such as stopping distance). Actually, planes have completely different tradeoffs from cars. For some perspective, see my post Why Don't Planes Obey Moore's Law? Speed has advantages in the air, and generally only worry about stopping distance on the runway.

Rob took a look at what is involved in autonomous driving, and how easy various scenarios are to handle:

  • Interstate 5: Main task is passing trucks, with about 10-20 objects to keep track of.
  • Highway 101: Main task is to get to your exit, with about 50-100 objects to keep track of, and lots of confounding factors, many of which are congestion related (too many cars, too little road, accidents).
  • 2nd Avenue (in Seattle in his pic, although it would apply in NYC, too): Over 200 objects to keep track of, with a lot of randomness (jaywalkers, cyclists, dogs, double-parked delivery trucks).
  • Parking lot: Main task is to find a spot, well over 200 objects to keep track of, with unclear rules, cars backing up, pedestrians, shopping carts, kids, and pets.

At each stage there are more objects to keep track of, increasing complexity of the task, and increasing number and complexity of challenges.

Is Driving Calvinball or Go?

But Robert had another distinction. Some problems are like Go, where the rules are known in advance and don't change, and the goal ("winning") is clear. This also has the effect that if you replay you get the same results.

 He drew a parallel with Calvinball, from the Calvin and Hobbes strip: "New rule, new rule. If you don't touch the 30-yard base wicket with the flag, you have to hop on one foot." The goal keeps changing and if you replay, you get different results.

This classification goes back to David Marr in 1977, over 40 years ago. Some problems are "clean" in that the problem has a known way of stating what the problem is and what the solution looks like (Go, Fourier transforms). Some problems are "messy" in that the problem is solved by simultaneous interaction of a large number of processes whose interaction is its own simplest description (Calvinball, protein folding). He went on:

The principle difficulty in AI is that one can never be sure whether the problem has a "clean" approach.

As an example, he had a picture taken from his hotel room window in Chicago:

The cars on the left are going around all the ride-hailing vehicles that are double and triple parked. Even the vehicles parked at the curb are parked illegally. But everyone manages just fine. Self-driving cars will need to manage this sort of situation, which is neither conforming to the rules, and involves a sort of negotiation between the drivers of all the different vehicles.

Here is the canonical stack for an autonomous vehicle:

But really it is much more complex since there are communication links, other vehicles, security, perhaps communicating infrastructure, and more.

 The Brakes Are On

The biggest design issues of all are that:

  • L3 systems use kW of power and fill the trunk. Don't think about putting luggage in the back of your autonomous taxi!
  • L4 and L5 need even more compute horsepower
  • Process node is slowing, node feature scaling is slowing, FinFETs run out of gas soon
  • Post-CMOS options range from "wildly improbably to impossible" said an unnamed fab engineer

The 3D language is attractive (but it is linear versus the old Moore's Law which was exponential). Here's imec's wonderful summary table:

Summary

Rob's conclusions:

  • Automotive is a key challenge and opportunity for the industry.
  • Safety, security, and resilience bring new challenges to machine learning.
  • Moore's Law matters and will influence what will be built.
  • 3D techniques provide a way forward.

NOVA

On Wednesday, October 23 , PBS's NOVA will feature an episode titled Look Who's Driving, about autonomous vehicles. I should give the caveat that I've not seen it, but the Computer History Museum thought it good enough to have a pre-screening and a panel discussion with industry experts and the producers, Menlo Park's Kikim Media. it is probably going to be on in other parts of the world (on Horizon in the UK, for example) although probably on a different date.

 

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.