DARPA: All Must Have Prizes

At CDNLive Boston, the invited keynote was by Dr. Bill Chappell of DARPA's Microsystems Technology Office. Since he took the position in mid-2014, he has focused the office on three key thrusts important to national security. These thrusts include ensuring unfettered use of the electromagnetic spectrum, building an alternative business model for acquiring advanced DoD electronics that feature built-in trust, and developing circuit architectures for next-generation machine learning.

DARPA was created the week after the Soviet launch of Sputnik in 1957, though it was then called just ARPA. The government's reaction to Sputnik was, "we never want to see that again." DARPA's mission was to ensure U.S. military technology would be more sophisticated than that of the nation's potential enemies. Over the years, DARPA has pioneered such technologies as packet switching (the Internet) and stealth.

The office operates in two ways. The primary approach (95-99%) are tradtional programs, commissioned by DARPA, put out to bid (normally) and managed by one of around 90 program managers. They're spending $2.9B in FY2016, on 250 programs run from 6 technology offices. Program managers serve for three to five years. As Bill put it, "when you get hired by DARPA, you know exactly when you are going to get fired."

He gave an example of one program: electronic brain interfacing. This started in 2010 with a quadraplegic controlling his prosthetic arm through an interface to his brain. In 2012, a woman who had had a brain implant for five years appeared on CBS's 60 Minutes controlling a flight simulator entirely by thinking.

The other way DARPA pushes technololgies is through challenges, with significant (million dollar) prizes for the winners. Perhaps the best known of these is the DARPA Grand Challenge to build an autonomous vehicle. For details on this, see my post 10 Years Ago Self-Driving Cars Couldn't Go Ten Miles. In 2004, none of the entrants finished. By 2005, five vehicles finished. In 2006, for the urban challenge, four finished. At that point, DARPA could walk away and leave it to the commercial sector to take the technology and run with it. There is simply no need to have another challenge with a million dollar prize to advance the state of the art.

Another famous challenge was the Red Balloon Challenge. Teams had to locate 10 red balloons placed around the United States and then report their findings to DARPA. Since teams couldn't be everywhere, they used social media to recruit other people. To DARPA's surprise, the contest was concluded in less than nine hours.

Some of the recent challenges around electronics were:

• autonomous vehicles
• robotics (enter a radioactive building with a bipedal robot and turn off a valve)
• cyber grand challenge (capture the flag)
• spectrum collaboration challenge (automatic share out radio spectrum without any pre-defined policy)

A lot of effort is currently going into machine learning on the basis that on a future battlefield, whoever has the best electronics is going to win. Until recently we used traditional programming with a human imparting how to do whatever needs to be done. But about five years ago there was a crossover and the big switch to training took place. Accuracy is up to about 99% today, which sounds good until you look at hundreds of millions of miles being driven daily. The move is towards doing the training without all the data being pre-identified, known as unsupervised learning. The table above shows the state of the art, with the green being where we are today and the pink, things we can't yet do. Another big challenge is explainable AI, since the military doesn't want unexplained events. Probably DARPA will move on from this area as the commercial world is picking up neural nets and deep learning in a big way.

In integrated circuits (ICs), the most interesting area to everyone at CDNLive of course, there is a focus on getting chips designed fast with limited resources. The program is called CRAFT. A couple of grad students over a summer designed the above chip, which is the most power-efficient convolutional neural network (CNN) chip for facial recognition. It is just 1.4mm2 in 40nm. In the past, DARPA funded a lot of EDA projects, but then they left it to companies like Cadence. This is the first EDA project funded for some time.

Heterogeneous integration is another important area, making it possible to mix commercial silicon with military silicon (or other exotic materials). The aim is to maintain control over the intellectual property (IP) while being able to manufacture anywhere. The above picture is a 65nm IBM base layer with chiplets in InP HBT chiplets and GaN HEMT chiplets.

So if you want to work with Bill's group (or DARPA in general), how does it work?

• Small efforts (<$2M) seedlings, can be any idea
• Bigger efforts (>$10M), DARPA comes up with the ideas and puts it out to open competition
• Funding of pre-competitive research between DARPA/Industry/universities. JUMP is forming now for 2018, the prior version was called STARnet.

DARPA also wants to be aware of what is already being worked on. "What is in your short window? We don't have to do that and, instead, can leverage it."

Next: Everything That's New About Ethernet

Previous: Electronic Design Automation Handbook