Chris Rowen: Tensilica’s rational trajectory
September 18th, 2012 by Peggy Aycinena
Chris Rowen is Founder and CTO of Tensilica, an IP company based in Silicon Valley. We spoke last week by phone to discuss how an IP company decides what and when to introduce new products.
I first asked to Chris for a brief history of the RISC [Reduced Instruction Set Computing] architecture he is closely associated with, and how that history segued into the founding of Tensilica.
From RISC to Tensilica …
Q: Can you give me a quick overview of the origins of RISC architecture?
Chris Rowen: RISC is a set of ideas that grew up in academia and IBM in response to increased architectures in both the mainframe and microprocessor worlds.
People saw machines with really high hardware costs being built for assembly [language applications]. However, as compiler technology got better, people said: If I want a compiler to run well, I don’t need fancy instructions. I only need a common set of instructions that run really fast. All other complex operations could be composed by the compiler out of these fast, simple operations.
RISC grew out of these compiler technology advances, and a recognition in the VLSI era that there was an opportunity to rethink the process of how the architecture could be put together. IBM’s John Cocke is mainly associated with these concepts, but similar ideas were emerging at the same time at Stanford with John Hennessy and at U.C. Berkeley with Dave Patterson.
They were all saying: Let’s try leveraging VLSI technology to improve how we design systems. Academics, in particular, realized universities could make important contributions here, and they quickly were showed dramatic results by developing microprocessors 5-to-10 times faster than anything currently in the market.
In 1984, John Hennessy pulled together a group at Stanford, including folks from Bell Labs and IBM, and other bright industry veterans coming from a number of places in the Valley. I was lucky to be part of the group that formed MIPS Computer Systems in September of that year, ironically to commercialize the Stanford MIPS Chip. We quickly decided, however, to move to a clean sheet of paper and define an entirely new chip.
MIPS has gone through a range of different business models since that time, starting in the chip business, moving into systems, and then on into microprocessor design. As MIPS’ products became wildly popular, the company was acquired by Silicon Graphics in 1992, and then spun out again in 1998.
Interestingly, because of the Silicon Graphics acquisition, MIPS became a pioneer in microprocessor licensing focusing primarily in designing for high-end devices. From 1994 to 1996, however, as ARM really got up and running, Silicon Graphics was largely absent from the processor licensing business. ARM, meanwhile, established itself as a leader in that space and executed well.
If MIPS had done something different, if we could roll back the tape and start again, the world would look very different today.
Q: Your Tensilica strategy emerged out of your MIPS experience?
Chris Rowen: Yes. In 1996, when I returned from working with Silicon Graphics in Switzerland, I decided to join Synopsys to run their Design Reuse group. They started in the USB and PCI interface logic area, but we tried to think much bigger.
It was clear that the centerpiece to IP-based design was always going to be the microprocessor; every chip will have a processors. It was going to be a very highly leveraged item for application-specific chip design.
We also recognized that the impact of Moore’s Law would be ever greater integration, and ever greater integration would mean ever greater specialization of the chip. Chips would look more and more like systems as they got bigger. I could see there were plenty of general purpose chips, but I thought there would be a market for more application-specific products.
Once I saw that, I knew there were big opportunities for these types of microprocessors. Nobody was doing it at the IP level, it was such a difficult technical problem. You had to work through all of the general questions of delivery, but also needed to narrow the application. I decided we needed to pursue broadly specific processors to attach a wide range of problems within a specific area.
Clearly to do so, we needed to solve the problem of getting good models cheaply while still maintaining a 5-to-10x improvement in efficiency. These results would outpace anything you could achieve through improved synthesis, faster transistors, better routers, or any other innovations within the silicon design world. And this was the motivation for Tensilica.
What we deliver is not the same as a general processor. We’re not competing head-on with MIPS or ARM, but are delivering application-specific processors. We’re meeting an increasing demand for good programmable and application-oriented architecture.
Q: How do you decide which application domains to go after?
Chris Rowen: That’s actually the beauty of the IP business model: If you make a chip, you take all of your [poker] chips and put them on one number. You must guarantee the one chip penetrates the market. With application specific IP, however, you don’t need to be so narrowly specialized. You’re only going to make pennies, or a dime at most, per customer’s chip that you’re involved with, so you are betting on a much wider set of numbers on the market.
In order to leverage all of that, we decided that instead of building application-specific cores, we would create tools that take high-level specifications and create customized processor RTL instantaneously, with high optimization and specialization. We also create the complete matching software tool chain. We therefore enable lots of different silicon designers to have the design of their dreams.
As Tensilica has matured, we’ve also discovered that design teams come in lots of different flavors. Some teams know exactly what application they want to do, which others only have general guidelines with respect to their goals.
For these types of customers, we have developed domains where we have a lot of in-house system knowledge. We can use our own generator for applications in audio, multimedia, baseband, and other areas. We generate processors in these various sectors, which in turn provide solutions and a starting point for many of our customers.
Q: In essence then, you’re providing design services?
Chris Rowen: No, not at all, because that implies having a team of engineers, interviewing the customer, getting specifications, and using a lot of sweat to deliver the specific thing the customer wants. We may go deep with a customer into the problem at hand, but we are not doing the design.
We’re relying instead on the automation we provide our customer for access to processor generation. We provide a set of software tools which generate standard products, not standard services, with a concentration on 3G, 4G, broadband, audio or video processors. True, we have a significant collection of cores that are already designed that can be further configured as needed by our customers. We establish the theme and the software base the customer then builds on.
A lot of people are not just looking for a processor they can program, they’re also looking for a collection of software and processors that work together nicely. We’re providing that, at a higher level of software and application-specific knowledge built into the processor.
Q: How do you decide what areas of specialization to concentrate on within Tensilica?
Chris Rowen: In every market, there are sub-markets, some of which are quite lucrative. Deciding which market or sub-market to pursue is a consequence of having this remarkable processor generator – our secret sauce for being nimble – and as a result, we are able to pursue emerging markets at a much faster rate than other IP providers.
Q: How do people know to come to Tensilica to meet their needs?
Chris Rowen: That’s a question at the core of our position as a corporation. Currently we have nearly two billion processors out there, so people have definitely heard of us. They’ve heard of our platform for automatic processor generation.
In markets where people are not as familiar with us, however, there are lots of design services providers who do know of us, as well as a rich community of system architects who continue to expect us to help them solve their problems. That’s our best opportunity to become known to their customers.
Of course, there are also people who have worked with us while at one company, who take their knowledge of Tensilica with them when they move onto another employer. They know we can provide a platform that’s far more flexible than what is traditionally expected in a hardware sense.
Q: How do you know when to expand beyond your existing expertise, versus a more conservative stance of sticking to your knitting?
Chris Rowen: That’s an excellent question. Naturally, we wouldn’t want to try to enter 10 new markets at the same time, even if we do have the appropriate technology. At the same time, there are some very clear lines that tell us it’s a good time to develop a new vertical application. We look at a number of criteria.
Number 1: We look for a domain which is a computationally hard problem where people need to do high throughput, or high performance, or high performance per watt, or high performance per dollar.
Number 2: There is a growing reason to move to a more programmable solution. In many of the markets where there are great opportunities, people have previously used hard-wired solutions but now seeing increased complexities in standards, be it video, audio, or baseband.
The customers in those cases say next time they will not go with a hard-wired solution, it’s too long of a design cycle and produces too many bugs. In that case, there is a compelling reason for them to go with something programmable, something that Tensilica can provide.
Number 3: Does this new technology sector fit into a larger market – the stuff of classic Marketing 101. Particularly as you only make pennies or dimes per unit, it’s important to justify the investment. We look at any market where there are lots of chips being designed,
Number 4: Is there a large customer base? What are our big customers doing? Are there other opportunities on a chip that’s already in production, but is now being revised by a new department within an existing customer? Every design we are involved with is a potential for expansion into other adjacent functions. We are always looking at those types of opportunities.
That combination of watching our customers, and watching high volume products in the market, is what convinces us it’s time to move into a new sector. Being nimble technically is built into our DNA, so when we see opportunities we’re able to react.
Q: You make it look easy. There’s an opportunity, so pursue it?
Chris Rowen: Of course, there’s always a complicated evaluation of opportunities that present themselves.
Do we have the knowledge in-house already, or is it readily available? And, a market we’re familiar with will outweigh a market we’re not as familiar with, and that’s often part of the debate we undertake before moving forward.
We also look at the risks: How certain can we be sure of a successful development? It may be an exciting trend in the industry, but are we able to embrace that trend?
And, how strong are the alternative solutions that already exist? Often the inertia of a pre-existing methodology already in place with the customer will win over a newer solution. Do we have the sales channels, and a technology that’s sufficiently superior to beat the inertia of the in-house solution.
We have to make a judgment call in considering that sort of competition and decide if we can beat it. As I said, the decision to move forward is only made after a very lively debate.
Tags: ARM, Chris Rowen, Dave Patterson, IBM, John Cocke, John Hennessy, MIPS Computer Systems, RISC, Silicon Graphics, Stanford, Synopsys, Tensilica, U.C. Berkeley