Tea Leaves at ICCAD: Struggling to see the cup half-full

Dreary sentiments notwithstanding from several panelists at an ICCAD evening session on November 2nd in San Jose, SRC’s Dr. Bill Joyner espoused optimism and energy for the future of EDA, even if said future doesn’t include the venerated Moore’s Law stretching off into infinity forever.

As moderator, Joyner convened the panel, “Moore’s Law is dying, EDA to the rescue!”, and turned over the podium straightaway to University of Pittsburgh’s Dr. Alex Jones for the first 20 minutes, which allowed the professor to report out on a 3-year Computing Community Consortium effort, just completed, to examine and exhume EDA from the doldrums.

The CCC’s group of 50+ academic and industry leaders have been meeting since 2012 at a series of SIGDA/CCC-funded workshops hoping to impact the future by nudging industry and academia into more productive avenues of research and development in design automation.

The report the committee published, “Workshops on Extreme Scale Design Automation (ESDA) Challenges and Opportunities for 2025 and Beyond”, was available in paper form at the back of the room during the ICCAD panel and has subsequently proved to be great reading, and fodder for a future blog. But this blog is a thumbnail sketch of the November 2nd discussion, so please read on.

******************
Moore’s Law is Dying, Long Live EDA!

The 90-minute panel, organized by IBM’s Frank Liu and handily moderated by Bill Joyner, included 6 PhDs: University of Michigan’s Todd Austin, GlobalFoundries Luigi Capodieci, Synopsys’ Patrick Groeneveld, University of Pittsburgh’s Alex Jones, IBM’s Leon Stok, and MIT’s Jacob White.

Joyner framed the discussion with the program intro: “After 5 decades of continuous progress, scaling has become increasingly difficult due to mounting technology challenges and prohibitive financial costs. The slowdown of scaling has brought big changes in the semiconductor industry landscape, as well as priority shifts in academic research funding. As a key player in enabling the benefits of Moore’s Law, EDA is also at a crossroads.”

As mentioned, Pittsburgh’s Alex Jones followed with a 20-minute analysis of the CCC report, his enthusiasm evident in one particular summary statement: “This uncertain and exciting environment is reminiscent of the beginnings of the EDA era!”

Michigan’s Todd Austin was next, with slides titled: The Good, Bad, and Ugly. He said there are enough good transistors for the foreseeable future, but it’s clear that at 22 nanometers, we’re talking about layers of no more than 10 atoms, a situation which clearly can’t continue to scale. Ergo, Moore’s Law will die.

Even worse, Austin said, “Bad Dennard Scaling is what we’re seeing on devices as they become faster and use less power, resulting in lackluster designs that can’t clock, and creating a gap between Moore’s law and where we want to be. The only thing we can do is to replace technology scaling with innovation.”

Austin turned to economics: “At 22 nanometers, it can take 70 million bucks to bring a chip to market!”

He noted that when a market’s too small to support such costs, the product simply won’t be produced. Hence, one way to save the industry is to lower the cost of bringing something to market.

To do that, Austin laid out a simple laundry list: We need to broaden the applicability of customized designs, explore designs based on approximate computing, have access to better tools to leverage the benefits of customization, embrace modular open-source IP, reduce verification costs using run-time repair and dynamic verification.

Austin closed with two distinct admonitions: “We’ve just got to embrace open source! And manufacturing is way too expensive! We can fix that by not giving silicon its personality in the fab [but using software instead to achieve] near-FPGA flexibility with near-ASIC optimization.”

GlobalFoudries’ Luigi Capodieci was next at the podium and began: “I subscribe to all of Todd’s vision [and because of his optimism], will back off of my despair over the poor vision for EDA. Yes, at 22 nanometers things are just a few atoms thick. However, at GlobalFoundries we will stop [production on a design] because of an issue down at the scale of .01 nanometers [so clearly we can deal with those geometries].

“The real problem is that the tools that are being employed have lost touch with what’s important in the design. In reality, we just pour down a big cement block and then, using EUV or a hammer, we open up the shape.

“Of course, EUV has now been delayed again until the next node. In the meantime, we have to make do at 20 nanometers with double, triple, quadruple patterning. [In reality, however,] you can actually do whatever you want to do except for one thing — the cost!”

Capodieci distinguished between two different types of gaps, the design gap and the patterning gap, and hinted that the latter is much more troubling. He cited the SPIE 2013 keynote from Dr. Chris Mack, which laid out three scenarios: 1) EUV enters production, and scaling continues; 2) 193-lambda lithography is combined with disruptive patterning, and scaling continues with the costs contained; 3) Both Moore’s Law and Dennard Scaling are redefined, geometric scaling ends, and the innovation and value moves up to the system level.

Either way, he said, we’re going to need collaboration between the foundry, EDA, IP, and design communities if we want to counter the current, despicable investment trend where companies like Snapchat get bazillions in funding, while EDA continues to get next-to-nothing.

Building on that pessimism, Synopsys’ Patrick Groeneveld stepped up to bat and proved that honesty is not always the best policy: “Moore’s Law is dying, long live EDA! I take this issue quite personally as I have been living with it my whole career. In fact, Moore’s law is the same age as I am.”

Groeneveld said it’s not really clear that things are slowing, and asked folks to look at Apple’s A4, A6 and A8 chips as evidence that scaling is chugging along quite nicely, thanks.

He then detailed his own analysis: “I have no private information here, but looked at Wikipedia data [and concluded] Moore’s Law is not slowing down at all. There’s no evidence of even a sputtering. In fact, Moore’s Law and Claasen’s Law are both alive! Yes, there is some slowing down, but there’s another 10 years until [we get to] 5 nanometers.”

Groeneveld then suggested that he was not the only one in the room who would be retired by then, so it should be less of a concern.

“And yes,” he continued, “there’s a lot more to be done, developing architectures that have better power/performance trade-offs for instance, but the EDA tools will compensate. At Synopsys, we will continue to offer the same quality [in our tools] and will make it all go a lot faster.

“We reduce inefficiencies in the design steps, [provide] better tool integration, and better tuning of the flow based on data mining. Moore’s Law is continuing because we can make it happen. It’s just no big deal!”

Groeneveld closed with a particular emphasis: “We clearly need to revitalize the relations between academic research and industry. The EDA industry should be doing more to help research work in academia, but we are not philanthropists. Somebody has to pay my salary!”

With that, IBM’s Leon Stok took the podium. “I don’t want to talk about death,” he began and got a big laugh from the audience. “But yes, it’s true that the productivity of chip design has fallen behind. Today, only Apple can pay for chips.”

Stok said the response in EDA should be an innovative reinvention by developing even better design reuse methodologies and embracing business models like EDaaS, Electronic Design as a Service. Stok also bragged on several research initiatives at IBM, and said the company’s success is based in part on “having an operating environment where things continuously work.”

The final at-bat went to MIT’s Jacob White, who channeled all of the cup-half-empty vibe in the room and none of the cup-half-full, by enumerating what’s wrong in the world and backed it up with what he warned would be a series of slides that would be confusing and fast, [and by inference, impossibly unfocused but brilliant].

White’s topics included: Kids are getting bad instruction today in robotics. We need to work on multi-technology initiatives. The key to improving EDA is to enrich online engineering education, whereby students will get answers back to questions they never knew to ask. Engineering education today, and electrical engineering and computer science education in particular, are so woefully lacking it’s hard to actually describe. Nobody listens when White tells them how bad it all is, and even worse, no one listens to him when he says he uses MATLAB. No one else seems to have heard of, and/or embraced MATLAB for design exploration. And finally, kids today are so stupid, they would rather work at Google than work at companies where real engineering is happening.

Yeah, it was pretty dark. But Bill Joyner re-established control at the podium, insisted on ending the evening on an optimistic note, and praised the value of software with this caveat: “Software makes it hard to make money. As soon as it gets specialized, you can’t make money.”

******************
Epilogue: Pop Quiz

Hopefully, after reading this brief account of the ICCAD panel on November 2, you are a confused as I was. The question of whether Moore’s Law, which is mortally wounded, can be saved by EDA was not really answered. Instead, the questions that were answered were as follows.

Q1) What comes after Moore’s Law-type scaling ends?
A1) It’s not clear, but perhaps really exciting things in cyber-physical systems, software, and bio-medical applications.

Q2) Why can’t Johnny read? Or design chips?
A2) Kids these days just aren’t very smart. Neither are their teachers or their academic institutions.

Q3) Should we care what’s going to happen in 2025?
A3) If you’re an academic, the answer’s yes. If you’re in EDA, the answer’s no.

Q4) Where do big foundries look for technical talent, given that nobody’s studying EE in school anymore?
A4) They look to grad students who know something about databases, that’s how bad it’s gotten.

Q5) Will EUV ever happen?
A5) Who the hell knows?

Q6) Why does innovation happen in Northern California with so much more joy than in New England?
A6) Answers may vary: i) Snow versus no snow; ii) They’ve got old stuff, old buildings, and can’t let go of The Academy thing; c) Dunkin’ Donuts versus Pete’s; d) It’s impossible to surf on the Atlantic Ocean, particularly on Cape Cod; e) Big Dig versus Golden Gate; f) Harvard/MIT versus Stanford/Cal.

Q7) Will there ever be a panel at ICCAD that reflects a demographic other than white guys?
A7) See A5.

Q8) Did you ever think somebody at an EDA conference would advocate aggressively for open source IP?
A8) No.

Q9) Do you ever think somebody at an EDA conference who’s not an academic will advocate aggressively for open source IP?
A9) No.

******************

Tags: ACM Sigda, Computing Community Consortium, Dr. Alex Jones, Dr. Jacob White, Dr. Leon Stok, Dr. Luigi Capodieci, Dr. Patrick Groeneveld, Dr. Todd Austin, Dr. William Joyner, EDaaS, GlobalFoundries, IBM, ICCAD, MIT, SRC, Synopsys, University of Michigan, University of Pittsburgh

This entry was posted on Thursday, November 13th, 2014 at 10:45 pm. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.