Betting On Power And Deep Learning

A very candid conversation with VC Jim Hogan about returns on invested capital.


Jim Hogan, managing partner of Vista Ventures, sat down with Semiconductor Engineering to talk about what investments deliver the biggest returns, how quickly, and why there are so few investors in some big growth areas. What follows are excerpts of that conversation.

23abe5fSE: What are you investing in these days and why?

Hogan: I have about 15 active deals right now. I generally invest in things that interest me and have four attributes. One, it needs to be a big market. Two, it has to be disruptive and serve the market in a sustainable way—or a way I can protect it, meaning I want patents around it. Third is what the team looks like. Sometimes I will invest in the team irrespective of the market. And fourth, I look at who else is investing. I’ve invested with people I shouldn’t have invested with and I learned a lesson the hard way.

I’m a semiconductor guy so that’s my focus. There has to be a pretty big IP portion of whatever I do. Tela Innovations, for example, is an IP company that we started about 10 years ago. I look for the ‘long poles in the tent’ of semiconductor design. Right now, it’s either energy—meaning power savings—or verification. Those are the two semiconductor areas I do outside of pure IP. Outside of semiconductors, it is generally some disruptive technology that allows us to use the things we learned in EDA, like simulation and synthesis, and increasingly deep learning. I have five to six of those type companies outside of semiconductors and EDA.

SE: Drilling down a bit, what are you looking for in IP that makes it interesting for you?

Hogan: Tela Innovations was pretty disruptive. We could see these limited design rules or structured design rules coming because of lithography. We created IP around that and found customers that were interested in us helping them. We anticipated that market because we got involved in a lot of DFM stuff. And, of course, I was involved in Artisan Components and those things that intersected at about 20nm. We had a pretty good idea that if we could intersect that point with some IP that we owned, we could license and make some money from it. Another example of IP is a company that I’m currently involved in that has been around a few years. We took a two-dimensional layout that we promoted in Tela Innovations and applied it outside of memory and digital to analog—it’s called Semiconductor Technologies Inc. in Utah. It’s probably doing $3.5 million to $4 million, so we invested in that. We’re likely to see some enterprise value growth as a result of their IP. At Tela we had a couple hundred patents. There are lots of patents around these things. In the most recent cases, capital investment has been $4M while the value of the companies is over $30 million to $40 million. It’s pretty efficient capital utilization.

SE: What’s the payback period?

Hogan: My average return is 8 years and my mean return is 7 years.

SE: So you are a patient investor?

Hogan: Everybody has to be. You talk to anybody that has been in this business for awhile and no one gets in and out in a couple of years. That is dot-comish/app stuff. With the hard technology it takes three years to prove the technology and another three years to validate the market. You are looking at a minimum of six to seven years before the company is worth anything.

SE: What types of companies are you looking at for your energy investments?

Hogan: There is going to be a lot of compliance for appliances—refrigerators, set-top boxes. The government is going to regulate how much a set-top box can draw in terms of power and energy. How can I anticipate that? We have a company down in southern California called Aggios that worked with the California Energy Commission, for example, to develop a hardware language and recommendations for energy compliance. We are calling it software-defined power. To make a big impact in energy you have two options. One is at the software-hardware signoffs through the bare metal layer. The other is down on the physical side at the transistor layer. I have a company in Denmark called Teklatech. They did some non-intuitive stuff. They figured out a way to look at dynamic voltage and optimize layouts for dynamic voltage by taking global constraints. While most guys only do certain sign-off on the budget, these guys took the approach of, ‘Let’s look at the energy requirement and place and route.’ They looked at it globally, loosened it up, and then shrunk the die. That was amazing and not intuitive for me. Then they really shrink the dynamic power. What you start to realize when you start talking about energy management—and this is true for Sonics, as well with its energy processing unit—if every transistor switches at the same time you get these huge peaks in dynamic voltage. Ideally what you want to do it shut off as much as the chip as possible and only turn on the things you need right away. What you can do is optimize turning things off and on and level the dynamic voltage requirement. In other words, you don’t throw all the switches at once. When people first started doing this, they turned everything on at once, had a big surge, and then it would come down. You can optimize this in several ways. You can do it like Aggios with software programmable energy at the bare metal/OS level. You can do it like Sonics is doing it with interconnect and the scheduling of events on the interconnect. Or you can do it like Teklatech, where you optimize the area by optimizing the dynamic voltage. Those things interest me. They’ve all taken relatively low money to get going and they are all profitable.

SE: These are really esoteric, inside baseball technologies. Is there enough of a market, and is it obvious as an investment platform?

Hogan: This really takes some knowledge, and you don’t see a lot of folks wandering into this area. I won’t say I have exclusively to myself, but generally speaking I see most deals that are going on. Then I can judge those deals based on my four critieria: protection, team, investment, and disruptive market. The other markets that I find myself in are the larger markets. These would be appealing to lots of investors. I have a company in Washington, D.C., involved with HIPAA-compliant open source called Blue Button. We are barely penetrating the market and we are already doing $6M. That tells you this market could be big. That takes good market and good computer science understanding.

SE: As you get into the more mainstream markets where people do understand that, what is the investment community and opportunity there versus what it would be in some of the more esoteric areas where you need deep working knowledge?

Hogan: If you look at EDA specifically, for example, the returns are between 2X to 7X. Once you get over 5X you are probably okay and comparable to other things. But it also takes a lot of time—seven years, which is outside the return parameters that an institutional fund can handle. Therefore, there’s not a lot of institutional money. Money usually comes in two ways—angel-type networks and strategic investors, like Intel used to do.

SE: As you look out at some of things exploding right now, like IoT, is that an opportunity or something you have to sit back and wait on to determine what we are really looking at?

Hogan: IoT means many things to many people. For me, it’s got three major components. It has the edge, where the physical world meets the digital world. This is typically analog sensors. There is the intermediate data collection and management, usually a mobile device, which is pretty well served at the moment. Smartphones are just getting better and cheaper every day, and that market is pretty much been parsed up and used. Is there opportunity there? Yes, there’s room for a new compression algorithm, but generally it’s too hard to get into that market. Then, there’s the cloud, which means big server farms. Then on the perimeter of the data center there’s opportunities to develop pre-processing capability for the servers to be utilized fully. There are opportunities there algorithmically. Hardware-wise, you see those increasingly serviced by FPGA outfits like Xilinx and Altera. I’m sure that is why Intel bought Altera. They recognized they needed to keep dominant in the server world. They had to ensure they had a programmable fabric to parse the daily terabytes of data they get. What do you do with all the data?

While we are still on the server topic, let me tell you about apps. The apps I’m interested in are cognitive apps, ones that take that data and do something with it. When you have a lot of data, deep learning starts making sense. You can teach a system how to find things and they can go out and work on it. But it needs a lot of data to really work well. As an example, I have an optics and signal processing background. I get together with some of my friends who have done this sort of thing with radiologists. We hooked up our technology capability to the lung cancer database. If you look at the MRIs, which are really high-fidelity images, there’s a lot of data. If you look at a MRI scan of your lungs, it is like 4 terabytes worth of data. The question was, ‘Can I train the system to be a radiologist?’ It turns out that you can, and it’s not really covered by patents all that much. Companies like Google and Intel are buying up all these software companies that are going to be horizontal in nature, in terms of providing a platform. The folks like me can develop apps on them. It’s pretty inexpensive to develop an app, so we did. We added a computer learning and optimization capability, got the data from the National Cancer Institute for nothing, put it in the cloud for $100 and ran a thousand servers on it for another $100. $200 later, we have the system telling us where cancer cells are. We taught the system to look for abnormal vascular growth, identify those, and then look for the termination of those things—and we found tumors. That’s where the world is going. We can rely on all the horizontal platforms that big guys, Google and Intel, are doing, and we can apply those applications pretty easily and utilize the cloud to store all the data. What’s that company worth? I don’t know. It will get sold in three to four years for maybe $10 million to $20 million. We won’t put much money into it. That is a pretty quick return with a pretty good ROI. There are a lot of those. That’s just a health example. In other vertical markets we are doing similar things. There is no way I want to be in the horizontal business as I don’t want to compete with Google and Intel. I want to find the vertical markets that can add some positive impact to mankind.

SE: So is this is the next big thing— being able to slice off pieces in narrow markets that can leverage what is already there?

Hogan: It is exactly like kids building phone apps. They don’t know how to program in assembly. They program in Python and build an app and sell it to Google for $20 million. There are going to be thousands of those things. I have this company in southern California, real esoteric stuff, called amorphous metal. We built a compound synthesizer. To give you an idea, the back of a cellphone, the stuff called liquid metal, that’s amorphous metal. The trick to it is that when you anneal this stuff, you never get to the boundary fractures, so as a result there are non-ferrous metals which are non-corrosive. It’s a huge savings. If you look at the Apple phone, for example, that’s titanium. It’s a lot cheaper if you can do it in a ferrous metal. We started working on that and then we built a compound synthesizer. The guy that did the Apple back, liquid metal guy in Pasadena, the best guy in the industry had five compounds he did in his 30-year career—Ph.D. We started building the synthesizer and thought, why don’t we build compounds everyday—teach the system how to build compounds. Remember, a guy in his whole career did 5 compounds. We did 50,000 a day, and it actually extends the patent for us, so we get this thing compounding itself at 50,000 claims a day on our patented technologies. It is amazing. That is deep learning plus the use of a cloud. Here’s the really interesting part. We noticed an anomaly in the curve for annealing, so we taught the system to find out what that means. The system found out that you can anneal aluminum differently. It was always done literally prior to that. We have a step function in aluminum annealing, and it created a whole new family of compounds. The system discovered something that humans would never see.

SE: The technology has gotten to the point now where technology is not the issue anymore—it is now a question of what can we do with it. That is where you are coming in?

Hogan: I’m a guy who will ask the right questions when given enough time. I’m never satisfied with just what people tell me—I have to ask the next question. Turns out that’s a pretty good skill to have right now.

SE: So as far as you are concerned, the technology race that you are seeing is just beginning?

Hogan: We are just at the beginning of the cusp. This is really going to accelerate. Look at the rate of patents—they’ve quadrupled in the last three to four years. People are just generating IP like crazy. I love going to the Maker Faire. It’s like Burning Man for geeks. You see all the crazy things people are doing. Some may be amateurish or toyish. But once in a while you think, ‘There’s a pretty good idea.’

SE: There’s a lot more of those good ideas coming out than in the past. We spent almost 50 years getting technology to the point where we can start doing these things.

Hogan: Exactly. In the late 1970s, when we were starting to move semiconductor design to digital forms and moving off mainframes to workstations, we felt like revolutionaries because we were giving the power of design to an individual. That individual could do what it had taken a mainframe a year or two to do. That changed everything. Raise that technology abstract by orders of magnitude, which is where we are today. We are right on the cusp of discovering some miraculous things.

SE: Looking back five years ago, did you feel the same way about investing then versus now? Were we on the cusp of things then or have we gotten much further along?

Hogan: Five years ago was 2011 and we were coming out of 2008-2009 downturn. I didn’t make any money then. Trying to get money to start something was impossible. The companies that got trapped in the 2008-2009 debacle consumed two more years of cash, and it wiped out a number of people. It was dire. It started going differently about two to three years ago when the cloud became more than just a curiosity. We started to see how we could use the cloud to solve interesting problems. We’ve shifted from outsourcing a data center to giving guys like me a tool that I can use to solve some really interesting problems. The Google folks might have realized it long ago, but practitioners like myself take longer. I’m not shy to use a tool once someone has developed it. Three years ago, guys like me, not just researchers, could start applying the cloud to problems.

SE: Is it still exit strategy for investment or is it about cash flow?

Hogan: For EDA there’s a threshold. You have to get a company to at least $10M before it is interesting to be sold, with some exceptions. It’s very difficult for a Cadence, Synopsys, Mentor, or Ansys to take a technology company and grow it. They have a channel, and anything that can go in their channel and be sold has multiplicity. It’s no surprise that we see exits to EDA companies—in case of Mentor 2X, Cadence 3X and sometimes 5X. That can go in a channel and immediately turn into revenue for the acquirer, so they can get an ROI in a couple of years.

Increasingly, companies are getting bought for IP. If you are at $5M or less, your company is going to be acquired for IP value. Sometimes you have a capability or an engine that is more strategic. A good example of that in the EDA world was Ansys’ acquisition of Gear. Gear had a big learning capability, and Ansys recognized it and used it across their entire product line. Now they have a deep learning capability. They were handsomely rewarded for that.

Second, you have to be profitable. These software companies are usually profitable at $5 million to $6 million per year and they start kicking off cash. They become cash machines, and it makes then attractive to acquirer. And then they don’t get forced to see for cheap.

Third, you have to get it done quickly. One of the problems with the 2008-2009 timeframe is that these companies took two or three more years to get to the point we are talking about. Everybody gets tired—the investors, the management team. They need to sell to get on with their lives. It’s helped that we’ve had cheap capital in post-recession.

SE: As interest rates go up does that change the picture?

Hogan: Yes, it changes the picture because as interest rates go up, people will move their cash to interest-bearing things with less risk. Our saving rates are really growing rapidly, probably a function of our aging population. I read some things on the development of the agrarian society. Survival of children went way up when the grandmother was available to help with nurturing the babies. That is true for companies. Survival goes up because us older folks still around and are still reasonably impactful, helping those companies improve their odds of survival and success. That’s true across the board.

Related Stories
Startup Corner
Semiconductor Engineering’s page devoted to startups
A Semiconductor Approach To Desalination
A new growth market for a very good cause.
Making Waves In Deep Learning
How deep learning applications will map onto a chip.
Neural Net Computing Explodes
Deep-pocket companies begin customizing this approach for specific applications—and spend huge amounts of money to acquire startups.
Power Options And Issues
From initial concepts to final sign-off, power has become one of the most challenging design problems.
MapD Makes GPUs First-Class Citizens
A San Francisco-based startup has come up with a novel solution for solving the big data analytics challenge by leveraging GPUs.
Silicon Catalyst: Semiconductor Incubator
With support from TSMC and Synopsys, new company funds and advises silicon startups.