Get email delivery of the Cadence blog featured here
I went to Bangalore to CDNLive India. It has a different structure from the other CDNLives that I've attended. It takes place over two days but almost nobody attends both days since the topics are different. Day 1 covered digital implementation, front-end, signoff and custom/analog. Day 2 covers PCB design, advanced verification methodology, formal, verification productivity and HW (Palladium/Protium).
This is not of interest to all readers but let me tell you about Indian visas. You have to apply online. I messed up my first application because i had to upload a scan of my passport, but I made the scan and never uploaded it, so my original application lapsed. Worse, I went to CDNLive Boston having left my passport on the scanner so I couldn't re-apply until I got back. I called the visa service Cadence uses and they said it was tight since it required three working days, and Monday was Labor Day. I was a bit suspicious of that, since it seemed unlikely that the US having a holiday would affect when a visa was approved in India. I reorganized my flights to enter India on Wednesday.
When I got back from Boston, I went online, filled out the form again. It said that it took 72 hours and should be available on Monday. There is a big difference between three working days and 72 hours when it is Thursday evening. My visa came through on Friday afternoon, about 18 hours later. So don't depend on it, but provided you are a straightforward case, you can probably get your visa at the last minute. And you need it to get on the plane, they won't let you board without something that looks like a visa approval.
When you get to Bengaluru, although I assume it is the same in other airports, don't follow everyone. There is a special desk to get the electronic visa processed, which takes about five minutes, and then they stamp your passport and you can bypass the long lines (mostly Indians returning home).
I hope that you know that I make a weekly video previewing the posts for the following week, called What's for Breakfast? In the hotel in India you could have breakfast from a wide variety of cuisines: American, Indian, Chinese, Japanese. The egg station, that normally just cooks fried eggs and omelettes in a hotel in the US, would also make to order masala dosa, a lentil pancake filled with spiced potato. Try getting that in your Denny's grand slam.
Last year, in the same dining room, we made the first ever What's for Breakfast? video. I think we've gotten better at doing them since.
Jaswinder Ahuja kicked off both days with a Cadence keynote. I won't go into details since I assume you already know the top-level Cadence messages around System Design Enablement. Like every other company in the semiconductor ecosystem, we are driven by the small number of fast growing segments:
The first day customer keynote was by Venugopal Puvvada of Qualcomm. For a more complete review, see Madhavi's India Circuit Blog post CDNLive India Keynote: Qualcomm on 5G and More. Venu is the VP Engineering at Qualcomm India. He told us all that it was fine not to put our phones on silent. "When a smartphone rings, I hear the ringing of a cash register for Qualcomm."
He talked a little about the difference between AI (artificial intelligence), ML (machine learning), and DL (deep learning). I think it is an uphill battle to keep the distinction, since these terms are tossed around somewhat interchangeably. AI is how to simulate intelligence with a machine. Machine learning is training machines to tell the difference between oranges and apples (cats and dogs seem to be popular, too). Deep learning is adding a lot more, to spot finer differences, whether the apple is from Kashmir or California (TIL they grow apples in Kashmir).
Venu pointed out that Qualcomm is one of the survivors in mobile. He joined ten years ago and since then many companies have dropped out (he was too polite to name them, but off the top of my head I can think of Texas Instruments, Freescale, NXP, ST-Ericsson, Intel). It is a ruthless industry with fast cycle times, consumer pricing, high complexity. What is needed for 5G, he said, is SoC expertise, world-class IP, EDA tools, advanced methodologies, and more. Qualcomm pushed power, performance, and area beyond what the process node delivers, needing better IP and better tools to get there.
Yield is really important. If you ship 100M devices at $10 then a 5% yield hit is $50M, which is real money. So it is essential to design so that worst case works for timing. Power is a little different since the voltage can be tweaked to match the silicon. The limiting factor is interconnect: RC as a percentage of delay keeps going up.
Finally, he got to deep learning, pointing out that we have largely stayed away despite enabling it for others. It is time for the VLSI industry to embrace ML/DL and look at area reduction, timing closure improvements, and other areas.
On the first day, the technology update was largely about Legato, the new Cadence memory solution, announced that morning. For more details see my post Legato: Smooth Memory Design.
During the day (and the second day) there were lots of breakout sessions on specific tools or techniques. I attended some of these and will write about them individually in the coming weeks.
The keynote the second day was by Dinakar Munagala of ThinCi. They are creating a real business out of AI, machine learing, deep learning. He thinks, and I don't disagree, that this is a paradigm shift (you have read Kuhn?) that will change a lot of things, not just autonomous driving, which is the high profile thing right now. ]
He thinks that the reason that this is all taking off right now is that we have more powerful computers (or al least a lot more of them), the algorithms have improved, and there are now big datassets around for training (I would say that imagenet alone changed things). This has meant that deep learning approaches have raced past the traditional approaches (find the edges, divide up the picture, and so on).
Outside of the obvious self-driving cars, he talked about drug discovery, and looking at which drugs will succeed. This may actually shrink the drug discovery process, by dropping drugs that will not be commercially successful early, but also make it possible to do personal drug discovery.
Most of the keynote was very general and I wondered what ThinCi actually did. They make a special-purpose processor for deep learning. It stacks up well against an NVIDIA p4 (even with the process-node difference, 28nm for ThinCi, 16nnm for NVIDIA).
Bottom line: AI is initiating a paradigm shift in computing.
Sign up for Sunday Brunch, the weekly Breakfast Bytes email