CASPA is the Chinese American Semiconductor Professional Association. Once a year they have their annual conference and dinner banquet. I ended up getting involved with them a few years ago when I stepped in with 24-hours' notice to moderate a panel session for them, plus they like to get some coverage. So I went along one Saturday recently to the Santa Clara Convention Center. Since it was just after Arm TechCon, I seemed to have been living there all week (in fact, the Arm free Wi-Fi was still running, so I made use of it; thank you Arm). The previous board of directors and officers transition to the new ones at this dinner. The new chairman, as of somewhere halfway through the salad, is Brandon Wang, whose day job is here at Cadence. He has agreed to an interview later in the year, so watch for that on Breakfast Bytes in a few weeks. The Conference The afternoon was taken up with the conference, with the theme being AI and Semiconductor Fusion: Best of Two Worlds. Wei Li, Intel Wei Li titled his presentation Accelerating AI from the Cloud to the Edge . He keeps being asked what Intel does for AI. Because he presented first, he got to put up his version of a slide that many other presenters had, showing the growth of data on the internet. The graph is a little odd since the units are gigabytes: a person generates 1.5GB/day, a smart hospital 3,000GB/day (aka 3TB/day), up to a smart factory generating 1 million GB/day which is one petabyte/day. AI of some sort is the most promising approach to pocessing that data and turning it into useful information for decision making. In this sense, AI is transformative, whatever the market segment, from health to finance, from government to industrial. He said that compute cycles for AI would grow 12X by 2020. Okay, that is only 3 years away, but still seems a really lowball prediction. Of course, Intel wants you to use Intel microprocessors for doing all that AI work. Since their share in datacenters is so large, any cloud training is likely to use a lot of Intel, even if there is an offload server of some sort. The Intel Nervana chip goes one step further and has separate pipelines for computer and the training itself, plus integrated high-bandwidth memory (HBM). Rob Aitken, Arm Next up was Rob Aitken. He gave a shortened version of the presentation he gave a couple of days before in the same building. I covered it in Greg Yeric and Rob Aitken Dive into the Details yesterday, so I won't repeat it all here. He did have a couple of examples that he didn't talk about during TechCon, such as a smart cordless drill. Why would you want a smart cordless drill? Well, first, since it will only work when your smartphone is nearby, it stops being an attractive item for thieves. But there are also things that a drill can do, such as controlling the torque precisely, that requires a user-interface which the smartphone delivers. Increasingly the question will be not what you might want to connect a given device, but why you wouldn't. Guna Ponnuvei, NVIDIA Since Guna works for NVIDIA, I was expecting him to talk about machine learning in autonomous vehicles. But he is responsible for bringup of chips and boards and his talk was titled Machine Learning in Advanced Automotive Quality . As NVIDIA has transitioned from largely selling GPUs to the gamer market, to selling GPUs for datacenter-based training and automotive chips, the requirements for reliability have gone up. I don't think these were real numbers, but he said that if a chip fails during manufacture, it costs $10. Once it has been assembled into a system, it costs $100. But if it fails once it is in a car, and the car has been sold, then replacing that box can cost $10,000. There is a lot to go wrong in a car. It can contain two miles of cables, and 100M lines of code (LOC). It also has a complex supply chain. A semiconductor company like NVIDIA or a DRAM company is what is called a tier-3 supplier. They sell chips to companies that assemble them into boards, often in China, known as tier-2s. The companies that the automotive companies deal with, such as Delphi, Bosch and Denso, are the tier-1s (and in automotive-land the automotive companies are called OEMs). Thee are a lot of challenges in this supply chain when problems occur since very little data is shared, meaning that it is hard to run statistical analysis. Plus, when a chip fails, the only way to identify it to the manufacturer is to remove it from the board and return it, which is expensive and slow. Guna wants "every chip to have a name" so that they can identify the component to tie it back to manufacturing data without having to rip it off the board. Doing this will make it easier for all the different contributors to the supply chain to use machine learning to optimize their piece of the puzzle. David White, Cadence David's talk was about how Cadence uses machine learning in our software and how we are productizing it. It was a similar presentation to the one from EDPS that I covered previously in EDPS: The Remains of the Day . The above diagram isn't really specific to EDA, although it covers what needs to be done, that is, splitting the machine learning into a training and development phase, and then using that trained data to get work done, in what David calls the "operational phase." This is often simply called inference, in the generic case. David listed the areas in EDA that he feels are the low-hanging fruit for using machine learning: Fast models for parasitic extraction Hotspot detection in layout Place & route Macro models for circuit simulation These largely have the pattern that there is a complex, computationally expensive algorithm. We can use machine learning to learn what that algorithm does, often under a limited input domain. Then, instead of using the expensive algorithm, we can use the trained model instead, which runs much faster. I talked about this in the context of drug discovery when I wrote about HotChips in Machine Learning for Higher Performance Machine Learning . Chris Rowen, Cognite Ventures Finally, the last session before the panel session, was Chris Rowen. His talk was a cut down version of his keynote from the Linley Processor Conferene that I covered in Rowen on Vision, Innovation, and the Deep Learning Explosion . Panel There was a panel session to wrap up the rest of the afternoon, moderated by Mario Morales of IDC. The panelists were: David Liu, CEO of PlusAI Jianxiong Xiao, CEO of AutoX XIngzhi Wen, CEO of Pinuitive Jeff David, CEO of StreamMosaic Alan Berezin of Lam Research Things were running late and dinner would be served on time, so the panel was a bit rushed. Things that were discussed were the time frame for autonomous vehicles (AutoX emphasized that they are working on solutions for 2-3 years out, one reason they ignore Lidar for now). StreamMosaic talked about who owned data since fabless companies don't own the foundry's data, even for their parts. Lam said that for their tools, the customer owns almost all the data, but the lack of standards is a barrier. PlusAI pointed out that a lot of the data in many areas like driving is generated and discarded, it isn't kept at all, so in that sense nobody owns it. Self-driving will be a feature of an increasing percentage of cars, but the cost has to get down to about $5K to make it widespread. Ajit Manocha Ajit is the relatively new CEO of SEMI, the association of the semiconductor manufacturing equipment and materials suppliers. He talked about Founders, Builders, and New Leaders Propel the Global Electronics Manufacturing Industry. He started by pointing out that it was his second time attending the CASPA banquet and it used to be about half the size (there were 1500 people there during the afternoon, fewer eating dinner itself because the room was full). The supply chain for semiconductor has done very well, with 20% growth from last year, which has not happened since 2000. He thinks it will be double digit growth for the next few years and semiconductor sales will pass $400B for the first time ever. There was $54B investment in fabs in 2016, and there are about a dozen companies that are in the billion dollar club. The industry is hot now, but the amount of new construction indicates that it will continue to be record breaking. Semiconductor seems to be a lot less cyclical than it used to be. It was two years of good growth and then three years of poor, but that era seems to have gone. For example, China is growing 25% per year, and has been doing that since 2001 (and by the rule of 70, if you grow at 25% per year, you double in size roughly every 3 years). Despite being Indian, Ajit has been to China many more times than he has been to India, they are growing in all areas: package and test, manufacture, and (especially) design. The growth in fabs in China (with about 24 under construction) is amazing, and it is not clear to me that they can all be successful. Fabs need process technology to run in them, and designs to fill them, just pumping money into their construction is not enough. I wouldn't be surprised if some of them turn out to be like those ghost towns with no people, buildings with no equipment inside. I won't try and tell you everything Ajit said and will focus on one area: bringing new young blood into the industry, which was the main theme he talked about. The industry has gone through three big stages. The founding of the semiconductor industry, from the 1950s to 1980s, with people like Andy Grove and Robert Noyce (and, of course, that guy with a Law). The growth era, from the 1980s to the 2000s. Now the new leaders, who largely remain to be identified, who will take semiconductors beyond CMOS and into adjacent areas. But it is not just new leaders that are required to take over the industry. Ajit wrapped up saying that: ...healthy kids do not want to go into STEM, but we need to make it cool. Our children’s lifestyle will be good, healthy and better. But will only happen if we as the young leaders sitting in this room bring people into the industry. Sign up for Sunday Brunch, the weekly Breakfast Bytes email.![]()
↧