In Silicon Valley, we take things like faster chips and Moore’s law for granted. It seems like progress is always inevitable, given the history of semiconductors and the accomplishments of chip makers such as Intel, which turns 50 this year.
But the fact is that advances in chip processing take hard work and smart people. This explains why Intel announced in April that it had hired Jim Keller as a senior vice president. He will lead the company’s silicon engineering efforts, which include putting together chips that can handle just about every task in a system.
Keller’s hiring has turned heads, as few can boast a trail of accomplishments like he has had as a chip architect, the person who defines what a whole family of chips could look like. Chip design is increasingly complicated, sometimes requiring thousands of engineers working on details that are as intricate as designing the flow of people and traffic in a 3D metropolis.
Linley Gwennap, longtime chip analyst at the Linley Group, speculates that Keller might be taking a fresh look at Intel’s aging x86 architecture from the ground up. Or he could be working on the next chip designs for artificial intelligence, or for marrying more than one chip in a system. Those all represent big ideas in the $412 billion semiconductor industry.
Keller isn’t saying what he’s working on yet, and maybe he won’t for years. But I spoke with him about his hiring and the publicity around his moves, which are watched carefully in the industry. He’s not a publicity seeker, but he has had a remarkable career as a chip architect. He started at Digital Equipment Corp., where he played a role in the 1990s design of the DEC Alpha processors.
In 1998, he moved to AMD, where he worked on the Athlon (K7) processor and led the K8 project that disrupted Intel’s 64-bit Itanium chips and gave AMD its first foothold in the lucrative market for server chips. Then, in 1999 — as the dotcom bubble was growing — he left for startup SiByte, which Broadcom acquired in 2000 for $2 billion in stock. When the bubble collapsed, so did the value of that deal and Broadcom’s own hypergrowth.
In 2004, Keller moved on to head engineering at P.A. Semi, a startup focused on mobile processors. Then he moved to Apple in early 2008. Apple also bought the P.A. Semi team, which went to work on the A series processors for iPhones. That was part of Steve Jobs’ strategy to become independent from chip makers, and it turned out to be a brilliant move that saved Apple billions of dollars.
In 2012, Keller sensed a shift coming again. Advances in PC processors were slowing. He rejoined AMD to lead a new microarchitecture, dubbed Zen. AMD launched the first chips based on Zen in 2017, and for the first time in years the company is rapidly gaining share on Intel. In 2015, Keller left AMD and joined Tesla to work on autopilot engineering for the company’s electric cars. (No doubt Tesla CEO Elon Musk is getting tired of paying Nvidia for AI chips for Tesla’s electric cars). Now, for the first time, this famous processor architect is working at the world’s biggest processor maker.
Here’s an edited transcript of our interview.
VentureBeat: I was looking at a Motley Fool story from last year that talked about this famous chip architect, Jim Keller. Apparently there can be a famous chip designer.
Jim Keller: I had no idea. I got to Tesla and they had a bio for me. I said, “Where did you get that?” They said, “It’s your Wikipedia page.” I thought, “I have a Wikipedia page?” I sent that to my mom. She was pretty impressed.
VentureBeat: It’s an interesting time. People have heard of things you’ve done. It’s interesting to me, having covered semiconductors since 1994. What do you think about this fame or notoriety? What does it mean about where we are?
Keller: I have no idea. Aside from bragging to my kids. I have a 12-year-old and a 13-year-old. They think it’s pretty cool that somebody called me a guru on a web page. You can’t beat that. But there are plenty of famous technologists. I’m not exactly a fame-seeker. I was surprised that there was a forum on some products I did, where people would passionately talk about what choices I made and why, what I was going to do next.
The fact that people are quite passionate about their computers — the gamer guys are especially passionate about the processor and the GPU and the next game and the framerate, and the people behind doing that. That’s kind of cool, that people are interested.
VentureBeat: I think they can tell that who’s behind something sometimes makes a big difference.
Keller: Yeah, that’s normal. The fact that they even go figure out that someone like me, who has relatively little public footprint — I’ve been lucky. I worked at Digital, which was a bloody great place to work. Then I went to AMD and some startups and Apple and back to AMD for a second try at that relationship. Tesla was super interesting. And then I got asked to join Intel, which has been great.
VentureBeat: It seems like you don’t mind bouncing around from job to job.
Keller: I really like working on the next product, the next really interesting problem. I’m an architect, but the problem of working with organizations to build products is a really interesting problem too. Finding the people — people are so interesting, right? There are innovative people. There are conscientious people. There are people who are great at teamwork. Verification people used to bug me, because I worried about everything. Then I realized that’s a genius plan. Some people worry about everything and you need that. But I like the dynamics of technical problems and organizing to go solve them.
VentureBeat: It seems like this becomes very strategic, every now and then. What you did at Apple helped give them independence from the chip vendors in a lot of ways. That seems like it helped them go more vertical and retain more of the profit in their business.
Keller: That was Steve Jobs’ idea, not mine. I’m a computer architect. When I was invited to work there, it was like, “Well, what am I going to work on?” They said, “That’s a secret.” So I hardly had an idea about that idea. Oddly enough, when I was there I worked with Intel on a couple of the Mac products, and that was interesting too. I’m the product guy.
VentureBeat: The notion that the big system companies can and should design their own chips, what do you generally think of that? Compared to the notion that chip-makers like Intel can supply that.
Keller: I’ve been in this business long enough to see both verticalization and horizontalization. The pendulum swings back and forth. It’ll continue to swing. There are different reasons for it at different times. We’re at a couple of inflection points in terms of technology change. Mobile is still running through the industry, low power. The expansion of cloud has been amazing. We’re in an AI revolution, if you go count the startups in that space.
Whenever you have inflection points, things move around a bit, but then there’s constant — foundries are really hard. High-end processor design is really hard. Bringing the components together to make a processor that’s really differentiated and really valuable is hard. If you look at the semiconductor market today, it’s continuing to grow on the top line. The pieces in the middle are moving around a bit. Some of it’s standard products from big companies and some of it’s customized silicon in-house. The mix of that has changed several times in the last three years, but the stability of — the super-hard problems tend to be designed by people who are experts at it.
VentureBeat: AI feels like it’s a new arms race going on right now, a fresh time of competition.
Keller: It’s fresh in the sense that a set of algorithms started really successfully solving these problems that have pretty general applicability. That’s the fascinating part. I know just enough about AI to know that — how does a neural network project information across these complex information plans, and how do we compute that? It’s an interesting problem set, and it turns out to be computationally intensive.
The style of that computation is a little different from the classic scalar computing or vector computing or graphics computing we were doing. It’s different enough. The application is broad. And of course there’s a little hype to make everything more exciting. Whenever there’s a change like this, especially when it’s visible from hardware all the way up to top-level software stacks, then a lot of people get engaged in that. Obviously Intel has been super engaged in that for a while. Most AI cycles still run on Intel. We’re making pretty big performance improvements on both the software and the hardware. That’s a super interesting problem.
VentureBeat: When you were outside of Intel, what did you think of Intel and what they needed?
Keller: I’d say it’s a little different. I’ve known Intel as a company, as a competitor and as a supplier, for a long time. I was curious to find out what the culture would be like. I have to admit, I’ve been in enough different companies that I try not to come in with too many preconceived notions about how things work, because everybody’s different.
The reason I came here is less to figure out what they needed or supply some ingredient than to participate in the coming computing change at the scale that Intel can do that. We all know the computing world is changing. It’s broadened out. The old mainframes disappeared. Then there were minicomputers, and then PCs and PC servers. Now we’ve gone up to real cloud computing. How long did cloud computing take to really happen? I still remember when IBM came out with the Grid. They didn’t know how to build it or sell it. That transformation took 20 years.
The mobile transformation is working its way through the whole ecosystem. The internet was super wild. AI is a similar scale. I’d say I came here to participate in that next wave of computing. But it wasn’t as if I thought Intel didn’t get it or something like that.
VentureBeat: It could be decades for this to play out as well, on the AI side?
Keller: Yeah, it’s definitely a big one. You can see it when you start seeing people coming out college writing completely different languages than they did four years ago. It’s the wave that’s going to roll through the computing world. AI and neural networks are upending science in a lot of ways. It’s going to be really interesting.
VentureBeat: Intel has a lot of resources now. You’ve been at other big companies, but maybe this is the biggest?
Keller: Yeah, Intel definitely has a lot of engineers. There’s no doubt about it. I quite like the culture. They have a culture of technical excellence and collaboration that’s kind of mind-boggling. I’ve been to lots of meetings where you want to work on a problem and you say, “Let’s get the experts together.” And then 50 people show up and they’re all pretty good. That’s fun.
VentureBeat: They could make you CEO.
Keller: [laughs] I doubt it! There are lots of other smart people here. I’m on the management committee. That staff is pretty good. It’s not as if any one person really sticks out.
VentureBeat: There are other things that could be done [at Intel]. It seems like a ground-up x86 might be a good idea, as well as all the work being done on AI chips. I don’t suppose you can share an opinion on that.
Keller: We have a great big Core line, and that’s pretty differentiated as far as performance and frequency. There’s a whole lot of innovation around data types and application sets that I think is interesting. The team building Atom for the smaller, compact core — I’ve been crawling through exactly where they’re at, and they’ve taken some big steps in recent years.
I don’t know that I look at it and think, “I need a from-scratch x86 core.” There’s a lot of methodology changes that are rolling through the industry as CAD development moves forward. I see it as there’s some methodology stuff that’s interesting to work on, and then there’s the application of some of the processors to new problems that’s interesting. And then whether we do a new core or rewrite something is more of a tactic than a strategy.
Strategically, how to execute what you want when you want — the methodology problem is the first order, and then what problems we solve, that’s shifting over time. The Intel guys already have a couple of pretty cool changes in the pipeline. As we evaluate all the different applications and what the customers are interested in, more things will come up.
VentureBeat: Are you hopeful about chip design in general, that it can still accomplish great things? There are always a lot of stories about Moore’s law coming to an end.
Keller: Yeah, of course. I was at an AI conference and someone asked me, “Is Moore’s law going to end?” They had all these reasons why it was going to end. I said, “I’ve been doing this for 35 years. Moore’s law has been coming to an end in the next five or 10 years that entire time.” I’ve decided not to believe that for the rest of my life. I’m not worried about Moore’s law.
It’s interesting to look back on what happened with silicon, when you go through these challenges. We didn’t really see planarized metal coming. But that really solved a big problem. Copper solved a big problem. Low-k dielectrics solved a big problem. Bigger radicals, 12-inch wafers — current fabs are built in sealed boxes, which is super cool, because then the clean room spec is way lower. Now EUV is coming up. FinFET transistors.
Intel has been a leader on quite a few of the big Moore’s law-level innovations. People say, “What’s happening? We’re running out of gas,” and we say, “Well, you have about a million people working every day who all believe in Moore’s law. They’re all collectively driving this, whether it’s lithography or chemistry or designs or metallurgy or packaging.” It’s a pretty broad palette of changes.
I’m not worried about Moore’s law. It’s gonna keep trucking. It’ll be bumpy. There will be years where a couple of clicks happen pretty smoothly and there will be other times when it’s a little more friction.
VentureBeat: I’ve mentioned that one of the fun things about chip design is that it’s not like designing car engines. You’re able to make a big difference sometimes.
Keller: Well, car engines have gotten a lot more efficient. The technology to design them and analyze them is super interesting. Of course, electric cars are going to make internal combustion engines obsolete, so that’s a different area.
Chip design is funny, because parts of it look just like what I did 30 years ago. And part of it is hilariously different. The first branch predictor I did was two kilobits of SRAM, and now, I don’t know, it’s 10 or 100 megabits. The scale of things changes radically. The number of transistors in a modern core is as many transistors as used to be in entire supercomputing centers. The scale difference is wild.
VentureBeat: Sometimes I like metaphors and their ability to explain things about chip design. Is there anything about where chips are now that lend themselves to a good metaphor that can help people understand something?
Keller: I don’t know. I’m looking for a metaphor. My motto has been bigger, better, faster, and smaller for the entire time.
VentureBeat: I think I remember something about doing your laundry in a washer and a dryer, and then each step that you take is a different part of the processing that happens.
Keller: Did I say that, or was that somebody else?
VentureBeat: I think I heard it at a chip conference a long time ago. It was an interesting notion about pipelines of chips. But sometimes that’s helpful to people in understanding things that have become so complex that they’re incomprehensible if you’re not involved in it.
Keller: Here’s the thing. Humans’ ability to understand things isn’t getting a lot deeper. So how do we make more complicated things? The most important thing is to build the abstraction layers at the right level. Over the years of computer design — you remember there was a debate about the wide person versus the deep, narrow person. Modern computer designers, we’re mostly specialists. For a while somebody could do the architecture, verification, physical design, and we called them a wide person. For the most part, on processor designs, people are specialists. We’ve cut the problem down that way.
Instead of building computers out of transistors, we have the transistor guys building transistors. The library guys make things like flip-flops and logic blocks. We have functional units. We build abstraction layers in both the technical skill level you’re operating at, and also at the design level. The big chips are parallel instances of lots of similar units. If you crawl through a computer, the number of devices is unimaginable. The state space the computer can exist in, with all the different programs running, is more complicated. But the actual span and scope of an engineer’s work — you keep re-factoring the different pieces of it in a way that people can have jobs that are interesting and challenging, but not so stressful that it’s impossible to get it done.
Modern computer designs are more complicated now than they were in the ’80s, but they’re way more predictable. There’s a certain grown-upness to the industry about how to go and build a project. That’s taken quite a while. There’s a craftsmanship to a lot of the steps. There’s art in the choices and art at the edges of new technology, but the actual building of the parts, it’s a craft with experienced practitioners. That dynamic is super interesting, if you’re a philosopher.
VentureBeat: As an architect, are you at the top of the pyramid, of those layers of abstraction? Are there fewer people doing what you do, and then further down there’s a lot more people?
Keller: I try to look across a lot of stuff, and what I see is that there are a lot of expertise domains where the experts there know more about it than I do. I’ve become a kind of generalist. If you thought of it as, there’s a pretty complex array of expertise that’s very deep — it’s not that hierarchical. There are independent things. There are software experts and floating point experts and memory architecture experts and branch predictor experts. They’re arrayed together and then we use some combination of organizational, tribal knowledge, and expertise to put that together.
I’ve had a long enough career and enough opportunities, I’m an expert in enough different areas, that I can sit down and talk turkey with people at a lot of different levels. But it doesn’t feel that hierarchical. There are hierarchical teams in terms of executing. We’re going to build this client part with these IPs and there’s a VP with a staff that goes and does that. But if you get into the technical level, you’ll find a fairly broad collaborative environment. Again, that dynamic is pretty interesting. Intel is quite well-organized that way.
VentureBeat: Does it feel like you’re designing the atomic bomb sometimes, when you’re in charge of organizing all this stuff, like the people who ran the Manhattan Project?
Keller: I don’t know. I used to joke — when I was at Digital we did very custom circuit design, and it was like building a wall. You start laying the bricks down, and then about halfway up you realize you want to change one of the bottom bricks. I felt more like a construction guy than an atomic bomb designer.
The scale of Intel is interesting because of all the different products and the talent and the people. But the technology is in so many real-world places — it doesn’t feel like atomic bomb technology. It feels more like a few thousand people working hard to build computers to make the world a better place. That seems pretty cool.
VentureBeat: Looking from the outside, I think we all expect some strategically important things to come from you.
Keller: [laughs] Well, we’ll see. I get involved in stuff, and at some point you look back and think, “Wow, that was a big move.” But while I’m doing it, it just seems like the next thing to do. The Apple story was just, “Let’s make the best possible phone chips we can.” That seemed great. Here we’re going to make the best possible server, the best possible client. We’re going to build great graphics. We’re participating heavily in the AI revolution. There’s a bunch of interesting problems there. We’re going to do some interesting stuff in that space.
The system architecture is really interesting, especially at Intel. We play in so many different components. We can architect across those boundaries without being beholden to somebody else’s problem or solution. And then at some point we’ll look back and say, “Hey, that was really interesting.” But to be honest, every single day we’re just coming into work and talking about stuff. My wife jokes that I’ve been drawing the same black diagram in my notebook for 30 years. She can’t tell the difference. Just two boxes with an arrow between them.