Post mortem: The inside story behind the collapse of chip startup Montalvo Systems

Montalvo Systems was one of the most ambitious chip startups in recent history and one of a few that dared to take on Intel in the low-power mobile processor market. Instead, it shut down in April, unable to raise a new round of financing.

The company had raised more than $84 million and hired an outstanding team of engineers in the hopes of grabbing a small piece of the $30 billion x86 (Intel-compatible) microprocessor industry. But it failed to get its chip out on time and sold its assets to Sun Microsystems for dimes on the dollar.

Understanding the company’s rise and fall offers lessons for entrepreneurs, especially those who would launch quixotic attacks against industry giants such as Intel. Clever design isn’t enough when it comes to getting a product out. For all its planning, Montalvo didn’t execute and was more than a year behind schedule, said several of the key figures involved in the company’s efforts.

The story of Montalvo also offers a peek peak into the game of chess that goes on between Silicon Valley’s biggest companies, and it raises a fatalistic question about whether the door is closed for venture-funded chip startups wanting to take on the world’s biggest chip maker, particularly in its flagship architecture, x86. That fatalism is a hard pill to swallow in optimistic Silicon Valley, where the belief in building better mousetraps rides high.

“x86 isn’t an easy thing to do,” said one of the key Montalvo figures. “There are 1,001 things that can go wrong with it.”

The Santa Clara, Calif. company was born in 2000 as Memory Logix in the garage of Peter Song, a microprocessor designer who, once upon a time, was an analyst for the Microprocessor Report. One of those articles was a prescient analysis that microprocessors would have multiple cores, or processors, on a single chip.

Song had figured out a way to create a low-power processor that was compatible with Intel’s chips. This was at a time when Intel hadn’t yet discovered its low-power religion. In an early interview, Song expressed intrigue in the possibility that Microsoft, having launched the Intel-compatible Xbox in 2001, might consider creating an x86-based portable gaming device.

Intel certainly gave no thought to Song’s company, which was working on low-power “synthesizable” cores. It was not nearly as ambitious as the later incarnation of the company would become.

But another rival, Transmeta, certainly awoke the sleeping giant. Transmeta, founded by former Sun Microsystems chip pioneer Dave Ditzel, raised hundreds of millions of dollars and managed to go public in 2000.

Transmeta’s plan to do a low-power chip was aimed at a chink in Intel’s armor. Intel’s own pre-Transmeta chips burned hot because, in its race against competitor Advanced Micro Devices, it had poured all of its engineering into chips that could run fast.

Transmeta focused on making chips for laptops and a new category of Internet-enabled handhelds. Intel, spurred by executives such as Pat Gelsinger, responded by placing a bigger premium on power efficiency instead of performance. The old tactics favoring performance hit the wall in May, 2004, when Intel canceled a chip that would have been the Pentium 5.

Intel’s designers in Israel came to the big company’s rescue. Led by Dadi Perlmutter, the Israeli team took the older Pentium III core and modified it for low-power mobile usage. In 2003, they launched the Pentium M, and Intel shipped it as part of a marketing campaign dubbed Centrino. It was a runaway success and crushed Transmeta, which had run into execution problems. Transmeta had to resort to licensing its low-power technology and suing Intel for patent-infringement. Intel also switched to doing dual-core processors on a single chip as a compromise between power and performance.

As Transmeta struggled, Song’s company drew attention. Vinod Dham, (pictured left) a former Intel engineering manager known as the “father of the Pentium,” was now running NewPath Ventures (and later NEA-IndoUS Ventures.) During 2004, he brought together a team that combined Song’s small group with two other veteran x86 architects, Mike Yamamura and Greg Favor. Both of them were behind much of the success at Intel rival NexGen, a 125-person company that AMD bought for $800 million.

Peter Glaskowsky, former editor in chief of the Microprocessor Report, joined as principal system architect. He suggested naming the company Montalvo Systems, after the Villa Montalvo arts center in Saratoga, Calif., where Transmeta debuted. As editor of the Microprocessor Report, Glaskowsky had a sharp wit and was always good for a quote. And he was frequently critical of Intel.

(Glaskowsky, pictured left, is tall at 6 feet 4 inches. I once walked through “The Star Trek Experience” with him at the Las Vegas Hilton during a trade show. He wanted to take a picture with Lt. Uhura, the actress Nichelle Nichols. I snapped the picture, but Glaskowsky was so tall that his head wasn’t in the shot.)

A former Broadcom engineer, Laurent Moll, joined as the SOC architect. Moll, Glaskowsky and Favor reported to Yamamura, who was vice president of engineering and for a time served as chief executive officer. They all came to work every day because Montalvo was the Big Show. If it grabbed just 1 percent of the x86 microprocessor market, that would amount to $300 million in sales. Everybody stood to get rich. The company’s internal ambition was to grab 20 percent of the notebook chip market.

Dham had a habit of swinging for the fences. He’d left Intel in 1995 and joined NexGen, where he met some engineers who would become key players at Montalvo. In 1998, he became CEO of a networking chip company, Silicon Spice, and scored big time when he sold it to Broadcom for $1.2 billion in 2000, at the height of the bubble. It was that sale that allowed him to make the move into venture capital investing.

Dham and his venture partner, Tushar Dave, wanted to give something back to their native India. So they decided to invest in companies that had big operations in India. They encouraged Montalvo to set up a large engineering team in Bangalore. This team would do the grunt work for Montalvo and in the process would create skilled engineers in India who could help the country migrate up the food chain in terms of high-tech prowess.

During 2004, the team debated its options. Dham wanted Montalvo to be more ambitious than some of the original engineers. One of his ideas was to fuse a processor and a graphics chip together. That’s an idea that AMD is pursuing. Some of the designers warned that the chip would become overly complicated if they did that, and the company abandoned the idea. It could pair its processor with cheap graphics chips instead, making its own chip easier to produce.

Montalvo also considered doing a chip that was only partially compatible with the Intel chips — a chip that would simply able to run Linux, not Windows that could be useful in the server market, but not in the consumer PC space. One of the engineers proposed doing a non-x86 cell phone chip, but that would have put Montalvo head-to-head with another giant, Texas Instruments. That idea was quickly and universally voted down.

Montalvo ultimately decided to create a Windows-compatible, four-core chip with asymmetric cores. It would include a couple of beefy cores to handle big jobs, but it also had a couple of smaller cores to handle easier tasks. The smaller cores would consume less power. Montalvo could also add some more chip functions that were normally used in the chip set, or a pair of chips that served as traffic cop inside a personal computer. If the plan worked, Montalvo would have the fastest performance laptop chip with the lowest power consumption and the lowest cost.

The whole scheme was one way to differentiate the chip from Intel’s product line, since at two points in its history, Intel had tried and abandoned efforts to throw a lot of different functions onto a single microprocessor. Montalvo filed more than a hundred patent applications.

Matt Perry, former chief of Transmeta, became CEO of Montalvo in 2005, and the company moved into the same building where Dham had his offices in Santa Clara, Calif. It later set up headquarters next door to Transmeta, just across the street from Dham’s offices. Dham invested in the startup and helped recruit a lot of talent. Money soon poured in by the tens of millions. Hiring engineers by the hundreds became the company’s biggest chore as it expanded in 2005 and 2006. In Silicon Valley, times were getting better and fewer engineers than expected joined the startup at the salary levels it could pay. Some engineers came from Transmeta, but for legal reasons, Montalvo had to avoid directly hiring Intel and AMD people.

The team in India had to handle more mundane but still critical tasks. At least half of the engineers had to characterize the behavior of x86 chips. That was a painful task. Essentially, someone had to buy a computer and run programs through it one at a time. Then the engineers would record the differences between the input and the output. They also had to come up with a verification suite to test all possible permutations of past applications. Ensuring x86 compatibility would require a ton of resources.

Montalvo clearly needed more money and engineers. Bangalore had become so hot that even Indian engineers weren’t cheap. In 2007, the company abandoned that idea of doing asymmetric cores. It decided it would be better off creating a high-performance laptop microprocessor with four identical cores and no chip set functions. This change didn’t cost a huge amount of time, but it would make a difference to investors later on.

The team now had to create a 65-nanometer chip that could beat Intel’s fastest laptop chips. It would essentially be a lot like AMD’s successful family of Opteron chips but designed for the mobile environment instead. But a return to the symmetric architecture was dangerous. It meant Montalvo would have to match or beat Intel’s low-power mobile processors in performance, since there was no other differentiation.

The team grew to about 280 or so people, with maybe 40 percent of the engineers in India. The chip fell behind schedule. Intel, meanwhile, launched its Core 2 processors in 2006 and became extremely competitive in mobile computers. With multiple engineering teams at its disposal, Intel was able to stagger its launch dates so that it had something big coming out every six months.

“Our goal had become performance, performance, performance,” said one Montalvo leader. “But the more we focused on that, the more dangerous it was when we fell behind schedule. Intel would easily close the gap with us.”

In the fall of 2007, it was clear there wouldn’t be enough money to finish the chip, dubbed the M-100. Montalvo started raising another round. It needed about $40 million to finish its design, or tape out. A total of $100 million would have helped the company get to high-volume manufacturing. But there was no guarantee that this redesigned chip was going to be on time and competitive against Intel, which was about to launch Centrino 2 (and still is). Centrino 2, coupled with new low-power processors coming, is likely to make Intel extremely competitive in the mobile processor market for some time to come, analysts say. Yamamura left Montalvo in 2007 for personal reasons and then eventually returned as an advisor to Redback.

Several venture capitalists said that the risks of delay were high and the consequences expensive. Theoretically, if the company were burning $3 million a month, then a one-year slip in the schedule would cost $36 million. A two-year slip would be $72 million. That’s a lot of money to cough up. And they you’re never sure if the product would be out of date by the time of its delayed introduction. On top of that, the company might have to spend a ton on marketing just to get the chip off the ground. Venture capitalists had to assume the company wouldn’t make money until its second-generation chip. If Montalvo succeeded with a second chip, it stood the chance of raking in billions of dollars.

Some of the technical people felt like the investors should have believed in the project more and understood that they were a year behind the most optimistic schedule, but they were really on the mark in terms of the realistic schedule, given the difficulties in hiring. The engineers felt that they got more done with fewer people than any other microprocessor startup in history. They felt they were nearly done with one of the top five most-sophisticated microprocessors ever created with an investment of perhaps 200 man years compared to 500 for other major chips. Even if it had raised new funds, Montalvo wouldn’t be spending more than what Intel, AMD, and IBM spent on their most ambitious chip designs.

Montalvo received remarkably little coverage from the mainstream press, perhaps because it was always in stealth mode. But there was a certain “prove it” mindset in the press when it came to challengers to Intel. Maybe if they grabbed that 1 percent of the x86 market, they would be worthy of coverage. Otherwise, who would know or care if startups like Montalvo had ever existed.

Yet Intel most assuredly was closing off all of the chinks in its armor. After former Intel CEO Craig Barrett spent billions on communications acquisitions, Intel took its eye off the core microprocessor business and left the door open for AMD’s Opteron in 2003. AMD got a second life, but it faced the wrath of Intel’s response. Paul Otellini took over as CEO, unloaded the communications businesses, and doubled down on the mainstay microprocessor business. He developed a “tick tock” business model of continuosly introducing new chips and manufacturing processes that stood a good chance of burying AMD and Montalvo along with it.

The money people were getting jittery in the spring of 2008. The economy was starting to teeter. It was looking risky. Some of the investors actually wanted to throw $150 million to $200 million at the company. But with the execution unclear, a key investor advised against pouring more into it.

The engineers felt like they were getting close and could finish the design in 2008. But CEO Perry had to start shopping Montalvo for an acquisition. He found that even big companies didn’t have the stomach for a fight with Intel. On March 31, a bunch of Montalvo engineers had lunch at the local hamburger joint. When they came back in the afternoon, they were out of work. Montalvo laid off two thirds of its employees. A few weeks later, the company closed its doors for good and Sun bought the assets. For Peter Song, that was the end of an eight-year journey. But Montalvo’s design efforts had really been under way for just about four years.

“Montalvo tried to do a little too much,” said one of the key engineers. “The reality is that most of the time, your first effort isn’t right. You have to go through another design cycle. That can take ten years.”

It is surprising that Montalvo didn’t find a savior that could keep it going. Jen-Hsun Huang, (pictured left) CEO of graphics chip maker Nvidia, recently harangued Intel for saying that the tasks performed by graphics chips would eventually be subsumed by Intel’s chips. No one would benefit more from owning a microprocessor design team than Nvidia. But, asked why he didn’t buy Montalvo, Huang said that he would rather create more interesting chips internally. (He didn’t offer a clue as to what those chips might be).

“You know me, Dean,” he said. “I think things through. Why would I want to do that? I could just hire their engineers.”

Nvidia only hired a tiny number of Montalvo people, including Laurent Moll, Montalvo’s SOC architect. As Montalvo ran out of cash, Sun Microsystems bought the assets (including patent applications) for an undisclosed price, but the Register reported it paid a little more than $5 million. (Apparently, it’s more than $10 million). Sun has hired more than 100 of the engineers, led by Perry. Sun got most of the local team and only a few of those in India. But some key talent hasn’t joined. Favor chose not to go to Sun.

Sun could use the technology to keep Intel at bay. The server maker designs its own Sparc chips but also buys from Intel and AMD. One of its intriguing options is to use Montalvo’s team to create a server chip with both Sparc and Intel cores on the same chip. That could give customers options for running a variety of software on their servers. Sun could also do its own x86 chip. But it could take a few years for such a design to materialize.

It remains to be seen whether anyone else will try to create an x86 startup to compete against Intel. But such an effort could take hundreds of millions of dollars, as Montalvo showed us. And it will only get more expensive over time. The target remains juicy, but elusive. Dham said only that it was a noble attempt.

Glaskowsky posted on his blog that he hopes to get permission to tell the story of Montalvo one day. So far, he hasn’t received it. Ever the Silicon Valley optimist, Glaskowsky believes someone will be able to compete against Intel and AMD in the next five years.

“If it was up to me, I would try again tomorrow,” he said. “I know we could do much better.”

0 comments