What Would Joe Do?
Peggy Aycinena is a freelance journalist and Editor of EDA Confidential at www.aycinena.com. She can be reached at peggy at aycinena dot com.
PDF’s Kibarian: Jedi Knight unplugged
July 23rd, 2015 by Peggy Aycinena
Luckily this week PDF Solutions announced the acquisition of Syntricity, a “provider of yield-improvement technologies and services for the IC process life cycle. Syntricity’s dataConductor platform is a comprehensive, enterprise-wide yield management system that leverages a thin-client architecture to provide a cloud-based SaaS or distributed enterprise solution, allowing users to access their data anytime, anywhere.”
That news gave me a chance to attend to a long-overdue task: Compose a blog based on a lengthy interview with PDF co-founder and CEO John Kibarian conducted in the lobby of the DoubleTree in San Jose earlier this year. On that day, Kibarian and I sat in the bar at the hotel, although sadly it was mid-afternoon and the bar was not yet open.
Instead, ours was an all-business conversation that gave me a chance to learn far more about the man, the enterprise, and PDF’s newest product release, Exensio, “an enterprise-wide, Big Data platform, which analyzes and reports critical data generated across the semiconductor ecosystem”.
I last interviewed Kibarian in 2005 and labeled him then a Jedi Knight. In the intervening years, little has changed. He’s still singularly focused on the technology, and his incredible obsession with the interminable analysis of manufacturing data. Nowadays there’s also a dollop of Haiku thrown into the narrative, however, which somehow adds additional weight to the whole Jedi discipline thing. It’s a discipline based on deep understanding, patience, intuition, and the ability to learn. Indefinitely.
WWJD – Let’s start with the requisite elevator pitch: What does PDF Solutions do? And may I remind you that you’ve said for years that PDF is not an EDA company, yet you’re currently serving as co-Chair of EDAC along with Lip-Bu Tan?
Kibarian [chuckling] – PDF’s business has been characterizing silicon. Over the last 10 years or so, we’ve been acquiring and/or developing systems for yield analysis.
Today, we’re analyzing up to 100 terabytes of data per year, a huge amount of semiconductor data from all of the major fabs, including FinFETs, 16, 14, and 10-nanometer technologies. These event nodes are very hard to control, with the interactions between design layout and manufacturing capability becoming bigger with every passing moment.
Now everything is collecting data – think IoT. Everything inside a factory has a computer or senor and is generating data. For the last 10-to-15 years, folks have had aspirations of using that data for improving production yield. Now using our data analytic tools and services, we’re helping our customers gain a lot of insights into how to make those improvements.
WWJD – Haven’t people always had the desire to bring manufacturing data upstream to enhance design effectiveness?
Kibarian – Yes, but [the vision is hard to implement]. We’ve seen cases where as much as a petabyte of data is being generated per week in a fab, yet none of the data is getting used for [tweaking] controls in the factory.
I’ve been all over the world in ‘alarm’ meetings when the yield in the fab has gone in a direction nobody likes. There’s been a crash and the factory manager has his hair on fire, and several thousand wafers have been thrown away.
In those cases, there’s always an attempt to see which signal has changed [in the manufacturing data], but it’s rarely just one parameter that tells you everything about the yield. It’s usually a combination of factors that impacts the outcome.
At PDF, we’ve always collected all kinds of manufacturing data. We thought that if you collect many different sources of data and combine them into a single data base, it can be used to decide which parameters are important, and which are not. And, of course, you need to have metrics by which to evaluate the parameters.
So starting in 2006, we tried to put these things together; we had a vision and went off to address it. Working from 2006 to 2009, we architected a completely different system, one that effectively dealt with Big Data analytics, with constantly larger volumes and faster input.
For instance, rapid thermal annealing – RTA – needs very high sample-rate collection. The resultant volume of data is gargantuan, literally terabytes of data collected per day. But that amount of data alone is not effective. You still need a physical model, and test vehicles and systems to help you apply some order to the [analysis].
[In that process], we looked at the database that Facebook developed, which is open and a combination of structured and non-structured data. Columns of numbers are structured, the photos unstructured. We set up our own system around the Cassandra database, using it to keep up with the amount and variety of [manufacturing] data being collected, and on top we built analytics and a visualization architecture. This last is really important, because if engineers can’t see it, they don’t believe it.
At the same time, we knew that every engineer wants to be able to write their own models in order to understand [the system they’re working on], so we also needed to support an analytics environment and things like the R and S+ statistical languages. [Kibarian’s familiarity with such things extends back to his PhD work at CMU.]
Eventually our work extended past 2009, and the bottom of the downturn, so by the end of 2011 we were able to do a pilot project with a leading customer in Japan. We had sign-off from them in 2012, and followed up with projects at other fabs.
Now today, we have a system in Exensio that we are very comfortable with and a number of successful installations under our belt. And we have found new applicability for our work.
Originally developed for logic fabs, we moved then to image sensors, and now our yield matching system is being used by analog devices manufacturers – all showing success using Exensio in the fab.
WWJD – Is PDF Solutions the only company providing this type of capability?
Kibarian – The industry used to make lots of decisions on quality metrics using engineering review boards. But now, as the amount of data goes up and orders of magnitude more chips are produced, people need more automated ways to make decisions. So this year, we’re bringing Exensio out to a broader customers base, both systems and fabless companies, collecting structured and unstructured data in real-time, and producing physical models based on our test-characterization data.
When we got into this, we were was just trying to make sense out of wafer data. But once you’ve built great technology, as we have, others start showing you other applications and other domains where the same kinds of systems apply. We’ve been really pleased with the response in the engineering community.
To be candid, we’re the only company that’s really made an investment in software and modeling for factories, not making big hardware for factories with software as an adjunct. We’re actually in the business of creating software that is the product. Certainly in terms of size, we are the largest company making those kinds of investments.
WWJD – Remind me of the timeline for the company.
Kibarian – We started in 1991 when I left CMU, moving from my post-doc right into founding PDF Solutions. In 1996, we moved to California where I did a second ‘post-doc’ in business, learning what was important to manufacturers, creating models for them, and learning business processes at the same time.
It was in 1996 that we really started focusing on our technology, Cadence got involved with us, and Lucio Lanza came on board. [Lanza has served on the Board of Directors since the mid-90s, and has been Chairman since 2004.] Then in 2001, we went public. There were only 10 companies at most that went public that year, but our business proposition and financials were strong enough [to support an IPO].
WWJD – Are there security concerns as PDF goes from manufacturer to manufacturer collecting data and analyzing it?
Kibarian – We’ve been very fortunate that our customers, our clients, have always had a high degree of confidence in our protecting their IP. It’s not lost on them that we’ve put huge investments into walling off our systems to guarantee customer security.
As a publicly listed company, we have every incentive to protect the IP of our customers. We can’t afford to have problems there, and our shareholders clearly understand the costs of guaranteeing security for customers are the table stakes.
In truth, our customers’ highest risks are in an employee walking out of their facilities with far more detail than we ever see. Good luck chasing those guys down.
WWJD – Could you ever have imagined the amount of data that’s being created today in all of this? Is this Big Data thing a blessing or a curse?
Kibarian – When we went into business, our ideas were about combining design information, consumable information and manufacturing data, and understanding the interaction between all of it. We didn’t realize at the time that we were working with Big Data, so it didn’t seem foreign to us.
It was in 2005/6 at a strategic planning meeting, that we concluded that since computing hardware had essentially become free, the problem was now all about the data – yet we still didn’t see it as strange.
We didn’t actually realize there had been a change [in attitudes toward Big Data] until we spent a lot of time in the field, saw how many companies run multiple, separate silos in their organizations, and how little data is shared across those silos. It was then that we started to [appreciate the relevance of our work] to Big Data.
WWJD – So Big Data is no longer news. It’s been embraced, and is producing results and business opportunities.
Kibarian – Yet despite the opportunities, the semiconductor industry is not attracting the best or brightest in North America right now. I was on GSA panel recently where participants were asked if they are excited about Moore’s Law. When I looked out to the audience, I recognized practically everyone who was there, [so few young people] are coming into the technology.
However, whenever folks [in one sector] get tired, there’s always a new crop that appears from new populations or new geographies.
For instance, I heard about the recent Maker Faire in the Bay Area on NPR and it sounded very exciting. Also, right after my GSA panel, I was in China for a different GSA event and found the enthusiasm there for all of this technology is totally different. For the people at Maker Faire and the engineers in China – these are places where you really see innovation.
And there are always new drivers in an industry. At one time it was the PC, and now it’s the mobile thing. I can’t fathom a guess as to what the next driver will be, but clearly today people are looking at the old way of following Moore’s Law and saying it can’t go on.
Yet companies like Intel and IBM continue to invest in R&D and advancements in technology, and they continue [to exhibit] wonderful returns on their investments.
WWJD – Is there an expectation for omniscience in the leadership of an organization?
Kibarian – In the tech business, leaders have to make decision that take 5-to-10 years to pay off. To do that, they need a deep understanding of the larger trends in the industry, how to position their business around those trends, and excellent intuition.
They also have to have confidence, patience, and the ability to learn based on a good understanding of the fundamentals. These are all skills one must have to build a successful business.
WWJD – You’ve always said PDF Solutions is not an EDA company, and not a capital equipment company. So what are you?
Kibarian – I’ve always seen us as a solutions company.
When I was an undergrad, I loved semiconductor manufacturing and concluded to you couldn’t [realistically hope] to start a chip manufacturing company. So I thought the next best way [to be involved] was to be in IP and the software needed to make manufacturing more effective. That’s where I focused in grad school, and where the co-founders of PDF and I focused our business.
Now, if you look at our financial model, it’s far closer to being an IP company first, a software-business model second, and a cap-equipment business-model third. Let’s look at these three models.
The part of the market we choose to serve doesn’t include a whole collection of companies. For that reason, it’s hard to be seen in Category One, an IP company.
That we affiliate with EDA companies may put us in Category Two, a software company, but EDA looks at us differently. We are an analytics company, not a software tools company. We are not in Category Two. Yet if I were running a chip company, anything that would help get our ideas into working silicon in the field could be seen as design automation. Nonetheless, EDA still looks at us as being different. We are not in Category Two.
In Category Three, we do not fit the model either. In the semiconductor industry, companies see value in IP and in manufacturing, and we [facilitate] both, but we do not make hardware. We do not fit the capital equipment model, and are not in Category Three.
The only solution is to invoke Haiku; 5-7-5 is the only solution to the problem. [Chuckling] Five parts IP, seven parts EDA, and five parts capital equipment.
WWJD – Something you said made me conclude that before 10 years ago you didn’t really have the capability to help companies bring silicon to market. Am I hearing correctly?
Kibarian – Fortunately for PDF, we’ve had a very long tenure in the industry. Usually Silicon Valley companies stop innovating after they go public, but we’ve built a very strong team. [We say to anyone], if you want to go to grad school and study manufacturing and characterization and the kinds of things that PDF does, we would be delighted to talk to you about joining our team.
In my own case, at the beginning of my career I thought about joining a semiconductor company. But in the end, they needed great marketing and great designs, but they didn’t seem to want to invest in the long-term.
But our business is about the long-term. If you’re someone like me who’s got a PhD in this field, you want to continue to be a necessary evil [in the business of semiconductor manufacturing]. We have a strong team, we’ve been together longer, and [pursued] bolder projects than others. The result has been tremendous effectiveness with our customers, and years of enabling their successes.
We can proudly point to what we’ve done for our customers in the past and what we’re doing today and say that it’s really great: Here are the issues that we have solved for you.
Our business has always been electrical characterization, and we continue to provide more data and information per unit of semiconductor manufactured than from any other source of data our customers utilize. Our data is important for design, for manufacturing control, and for the business of our customers. Our business model is about understanding our customers’ models faster, and giving them more info on those models than anything ever before. We have created terabytes of data for our customers, and helped their engineers make better decisions as a result.
WWJD – If I engage with PDF Solutions, how will my manufacturing and yield improve?
Kibarian – We will deploy our software systems, connecting to many different data sources across the process. You’ll then build a model about how each [tool in the manufacturing flow] should behave, and create a relationship between tool behavior and yield. When you have enough data to support your ability to understand that relationship, you’ll react. You’ll move quickly to [optimize] your manufacturing capacity and yield.
It’s true that it’s hard to code judgment, to take the human out of the [decision-making] equation. Cognition is difficult to code. Yet everything from slurry manufacturing to a shift in a batch of wafers, these small thing can impact output and yield. What we’re trying to do is to make the humans in [the flow] more effective. In the end, the data and the human judgment are both important.
You’ll never be at 100-percent [in your decision-making], but if the info is there along with systems that can help you model with the data, you’ll be closer to putting [a solution] together.
WWJD – Is more data always better?
Kibarian – The answer is always yes for a data junkie, who would always rather have more info than less. The last thing you want to do is find yourself overlooking a problem because you didn’t have the [relevant] data.
WWJD – What compute platform comes after CMOS, not to bring up something that might put PDF out of business with its current emphasis.
Kibarian – I’m always trying to think in my own mind what phenomenon we are trying to explain, and I keep coming back to how much energy does it take to do information processing, the moving around of data and the computation.
I tend to think there will eventually be a different type of processing element, potentially even binomial computers. The bipolar era drove us through the 1970s and 80s, and then computing moved in to CMOS.
Bipolar didn’t die, however. It continued on in power and other applications, although it was no longer a driver. So even if you ask about a computing element beyond CMOS, CMOS will still remain another element [in the constellation of choices].
WWJD – Nonetheless, is it time to look for new computing platforms?
Kibarian – Certainly the human brain is a magnificent piece of computing, and possibly a goal, but we’re nowhere near that potential with our hardware.
WWJD – Has disaggregation in the semiconductor industry helped PDF succeed?
Kibarian – Yes.
WWJD – Why not allow PDF to be purchased by a fab or semi company, allow your expertise to become the ‘secret sauce’ for one particular company?
Kibarian – By providing what we do, information for improving systems, we are able to deploy across all facilities in the world. Our systems then become richer in their capacities, and more capable. If you were to limit that [process] to just one manufacturing facility in the world, we would lose our potential to learn and grow.
20 July 2015 – PDF Solutions Inc. announced today it has completed the acquisition of Syntricity, Inc., the industry-leading hosted solution for characterization and yield management. Syntricity’s dataConductor platform is a comprehensive, enterprise-wide yield management system that leverages a thin-client architecture to provide a cloud-based SaaS or a distributed enterprise solution, allowing users to access their data anytime, anywhere.
dataConductor has been in use by many leading global fabless semiconductor companies for almost two decades. This acquisition extends PDF’s reach to those customers who prefer a SaaS solution. With the addition of dataConductor to PDF Solutions’ Exensio platform, Exensio now manages and analyzes over one petabyte of manufacturing data from PDF Solutions’ semiconductor customer base. PDF has a successful history of hosting customer data for characterization vehicles (CVs). Adding Syntricity’s hosted product data solution is a powerful extension.
Syntricity’s technology is highly synergistic with PDF Solutions’ product base and will be integrated into the Exensio platform, which includes the Exensio-Yield, Exensio-Control and Exensio-Test modules.
Key dataConductor features include:
Steven Griffith, CEO at Syntricity, is quoted: “We are thrilled to join the PDF Solutions team, working together to enter new markets and accelerate the development of innovative yield management solutions for our customers. Our combined customer base represents the lion’s share of companies in the IDM and fabless markets. We look forward to accelerating our development plans, establishing more collaborative relationships with our customers, and exceeding their expectations with a comprehensive set of solutions.”
3 Mar 2015 – PDF Solutions Inc. announced today the release of its state-of-the art solution focused on semiconductor yield maximization. Exensio, an enterprise-wide, Big Data platform, analyzes and reports critical data generated across the semiconductor ecosystem. The solution also provides tools and immediate information for foundries to rapidly adjust equipment for higher yields and reduced variability. Exensio provides a competitive edge to semiconductor and fabless companies, as well as foundries, by providing actionable information to rapidly identify and resolve manufacturing issues.
Inevitably, process complexity continuously increases to keep up with today’s advanced integrated circuits, however, there is an exponentially growing amount of data that needs to be analyzed. Most of today’s yield analysis methodologies and infrastructures are not adequate for finding root cause unless they have ways to effectively process Big Data. Additionally, manufacturing issues can result in high variability and low product yields, causing delays in product launches, as well as reduced revenue and profitability. Using Exensio, foundries and IC manufacturers can provide measureable value and faster time to market to their customers.
John Kibarian, President and CEO at PDF Solutions, is qutoed: “PDF provides our customers with solutions and services to proactively identify and understand the impact and relationship amongst the extremely complex factors that drive yield in today’s manufacturing environments. The leading-edge technology in Exensio helps our customers quickly identify and address product and process issues, which result in increased wafer yields. The Exensio unified analysis infrastructure enables our customers to rapidly resolve critical root cause relationships and leverage these relationships with predictive actions.”