Learning from Failure

September 25, 2006

Learning from failure is a hallmark of the technology business. Nick Baker, a 37-year-old system architect at Microsoft, knows that well. A British transplant at the software giant's Silicon Valley campus, he went from failed project to failed project in his career. He worked on such dogs as Apple Computer's defunct video card business, 3DO's failed game consoles, a chip startup that screwed up a deal with Nintendo, the never-successful WebTV and Microsoft's canceled Ultimate TV satellite TV recorder.

But Baker finally has a hot seller with the Xbox 360, Microsoft's video game console launched worldwide last holiday season. The adventure on which he embarked four years ago would ultimately prove that failure is often the best teacher. His new gig would once again provide copious evidence that flexibility and understanding of detailed customer needs will beat a rigid business model every time. And so far the score is Xbox 360, one, and the delayed PlayStation 3, nothing.

The Xbox 360 console is Microsoft's living room Trojan horse, purchased as a game box but capable of so much more in the realm of digital entertainment in the living room. Since the day after Microsoft terminated the Ultimate TV box in February 2002, Baker has been working on the Xbox 360 silicon architecture team at Microsoft's campus in Mountain View, CA. He is one of the 3DO survivors who now gets a shot at revenge against the Japanese companies that vanquished his old firm.

"It feels good," says Baker. "I can play it at home with the kids. It's family-friendly, and I don't have to play on the Nintendo anymore."

Baker is one of the people behind the scenes who pulled together the Xbox 360 console by engineering some of the most complicated chips ever designed for a consumer entertainment device. The team labored for years and made critical decisions that enabled Microsoft to beat Sony and Nintendo to market with a new box, despite a late start with the Xbox in the previous product cycle. Their story, captured here and in a forthcoming book by the author of this article, illustrates the ups and downs in any big project.

When Baker and his pal Jeff Andrews joined games programmer Mike Abrash in early 2002, they had clear marching orders. Their bosses — Microsoft CEO Steve Ballmer, at the top of Microsoft; Robbie Bach, running the Xbox division; Xbox hardware chief Todd Holmdahl; Greg Gibson, for Xbox 360 system architecture; and silicon chief Larry Yang — all dictated what Microsoft needed this time around.

They couldn't be late. They had to make hardware that could become much cheaper over time and they had to pack as much performance into a game console as they could without overheating the box.

Trinity Taken

The group of silicon engineers started first among the 2,000 people in the Xbox division on a project that Baker had code-named Trinity. But they couldn't use that name, because someone else at Microsoft had taken it. So they named it Xenon, for the colorless and odorless gas, because it sounded cool enough. Their first order of business was to study computing architectures, from those of the best supercomputers to those of the most power-efficient portable gadgets. Although Microsoft had chosen Intel and NVIDIA to make the chips for the original Xbox the first time around, the engineers now talked to a broad spectrum of semiconductor makers.

"For us, 2002 was about understanding what the technology could do," says Greg Gibson, system designer.

Sony teamed up with IBM and Toshiba to create a full-custom microprocessor from the ground up. They planned to spend $400 million developing the cell architecture and even more fabricating the chips. Microsoft didn't have the time or the chip engineers to match the effort on that scale, but Todd Holmdahl and Larry Yang saw a chance to beat Sony. They could marshal a host of virtual resources and create a semicustom design that combined both off-the-shelf technology and their own ideas for game hardware. Microsoft would lead the integration of the hardware, own the intellectual property, set the cost-reduction schedules, and manage its vendors closely.

They believed this approach would get them to market by 2005, which was when they estimated Sony would be ready with the PlayStation 3. (As it turned out, Microsoft's dreams were answered when Sony, in March, postponed the PlayStation 3 launch until November.)

More important, using an IP ownership strategy with the chips could dramatically cut Microsoft's costs on the original Xbox. Microsoft had lost an estimated $3.7 billion over four years, or roughly a whopping $168 per box. By cutting costs, Microsoft could erase a lot of red ink.

Balanced Design

Baker and Andrews quickly decided they wanted to create a balanced design, trading off power efficiency and performance. So they envisioned a multicore microprocessor, one with as many as 16 cores — or miniprocessors — on one chip. They wanted a graphics chip with 60 shaders, or parallel processors for rendering distinct features in graphic animations.

Laura Fryer, manager of the Xbox Advanced Technology Group in Redmond, WA, solicited feedback on the new microprocessor. She said game developers were wary of managing multiple software threads associated with multiple cores, because the switch created a juggling task they didn't have to do on the original Xbox or the PC. But they appreciated the power efficiency and added performance they could get.

Microsoft's current vendors, Intel and NVIDIA, didn't like the idea that Microsoft would own the IP they created. For Intel, allowing Microsoft to take the x86 design to another manufacturer was as troubling as signing away the rights to Windows would be to Microsoft. NVIDIA was willing to do the work, but if it had to deviate from its road map for PC graphics chips in order to tailor a chip for a game box, then it wanted to get paid for it. Microsoft didn't want to pay that high a price. "It wasn't a good deal," says Jen Hsun-Huang, CEO of NVIDIA. Microsoft had also been through a painful arbitration on pricing for the original Xbox graphics chips.

IBM, on the other hand, had started a chip engineering services business and was perfectly willing to customize a PowerPC design for Microsoft, says Jim Comfort, an IBM vice president. At first IBM didn't believe that Microsoft wanted to work together, given a history of rancor dating back to the DOS and OS/2 operating systems in the 1980s. Moreover, IBM was working for Microsoft rivals Sony and Nintendo. But Microsoft pressed IBM for its views on multicore chips and discovered that Big Blue was ahead of Intel in thinking about these kinds of designs.

When Bill Adamec, a Microsoft program manager, traveled to IBM's chip design campus in Rochester, NY, he did a double take when he arrived at the meeting room where 26 engineers were waiting for him. Although IBM had reservations about Microsoft's schedule, the company was clearly serious.

Meanwhile, ATI Technologies assigned a small team to conceive a proposal for a game console graphics chip. Instead of pulling out a derivative of a PC graphics chip, ATI's engineers decided to design a brand-new console graphics chip that relied on embedded memory to feed a lot data to the graphics chip while keeping the main data pathway clear of traffic — critical for avoiding bottlenecks that would slow down the system.

Stomaching IBM

By the fall of 2002, Microsoft's chip architects decided they favored the IBM and ATI solutions. They met with Ballmer and Gates, who wanted to be involved in the critical design decisions at an early juncture. Larry Yang recalls, "We asked them if they could stomach a relationship with IBM." Their affirmative answer pleased the team.

By early 2003, the list of potential chip suppliers had been narrowed down. At that point, Robbie Bach, the chief Xbox officer, took his team to a retreat at the Salish Lodge, on the edge of Washington's beautiful Snoqualmie Falls, made famous by the "Twin Peaks" television show. The team hashed out a battle plan. They would own the IP for silicon that could take the costs of the box down quickly. They would launch the box in 2005 at the same time as Sony would launch its box, or even earlier. The last time, Sony had had a 20-month head start with the PlayStation 2. By the time Microsoft sold its first 1.4 million Xboxes, Sony had sold more than 25 million PlayStation 2s.

Those goals fit well with the choice of IBM and ATI for the two pieces of silicon that would account for more than half the cost of the box. Each chip supplier moved forward, based on a "statement of work," but Gibson kept his options open, and it would be months before the team finalized a contract. Both IBM and ATI could pull blocks of IP from their existing products and reuse them in the Microsoft chips. Engineering teams from both companies began working on joint projects such as the data pathway that connected the chips. ATI had to make contingency plans, in case Microsoft chose Intel over IBM, and IBM also had to consider the possibility that Microsoft might choose NVIDIA.

Hacking Embarrassment

Through the summer, Microsoft executives and marketers created detailed plans for the console launch. They decided to build security into the microprocessor to prevent hacking, which had proved to be a major embarrassment on the original Xbox. Marketers such as David Reid all but demanded that Microsoft try to develop the new machine in a way that would allow the games for the original Xbox to run on it. So-called backward compatibility wasn't necessarily exploited by customers, but it was a big factor in deciding which box to buy. And Bach insisted that Microsoft had to make gains in Japan and Europe by launching in those regions at the same time as in North America.

For a period in July 2003, Bob Feldstein, the ATI vice president in charge of the Xenon graphics chip, thought NVIDIA had won the deal, but in August Microsoft signed a deal with ATI and announced it to the world. The ATI chip would have 48 shaders, or processors that would handle the nuances of color shading and surface features on graphics objects, and would come with 10 Mbytes of embedded memory.

IBM followed with a contract signing a month later. The deal was more complicated than ATI's, because Microsoft had negotiated the right to take the IBM design and have it manufactured in an IBM-licensed foundry being built by contract chip maker Chartered Semiconductor. The chip would have three cores and run at 3.2 GHz. It was a little short of the 3.5 GHz that IBM had originally pitched, but it wasn't off by much.

By October 2003, the entire Xenon team had made its pitch to Gates and Ballmer. They faced some tough questions. Gates wanted to know if there was any chance the box would run the complete Windows operating system. The top executives ended up giving the green light to Xenon without a Windows version.

The ranks of Microsoft's hardware team swelled to more than 200, with half of the team members working on silicon integration. Many of these people were like Baker and Andrews, stragglers who had come from failed projects such as 3DO and WebTV. About 10 engineers worked on "Ana," a Microsoft video encoder chip, while others managed the schedule and cost reduction with IBM and ATI. Others supported suppliers, such as Silicon Integrated Systems, the supplier of the "south bridge," the communications and input/output chip. The rest of the team helped handle relationships with vendors for the other 1,700 parts in the game console.

Ilan Spillinger headed the IBM chip program, which carried the code name Waternoose, after the spiderlike creature from the film "Monsters, Inc." He supervised IBM's chief engineer, Dave Shippy, and worked closely with Microsoft's Andrews on every aspect of the design program.

Games at Center

Everything happened in parallel. For much of 2003, a team of industrial designers created the look and feel of the box. They tested the design on gamers, and the feedback suggested that the design seemed like something either Apple or Sony had created. The marketing team decided to call the machine the Xbox 360, because it put the gamer at the center. A small software team led by Tracy Sharp developed the operating system in Redmond. Microsoft started investing heavily in games. By February 2004, Microsoft sent out the first kits to game developers for making games on Apple Macintosh G5 computers. And in early 2004, Greg Gibson's evaluation team began testing subsystems to make sure they would all work together when the final design came together.

IBM assigned 421 engineers from six or seven sites to the project, which was a proving ground for its design services business. The effort paid off, with an early test chip that came out in August 2004. With that chip, Microsoft was able to begin debugging the operating system. ATI taped out its first design in September 2004, and IBM taped out its full chip in October 2004. Both chips ran game code early on, which was good, considering that it's very hard to get chips working at all when they first come out of the factory.

IBM executed without many setbacks. As it revised the chip, it fixed bugs with two revisions of the chip's layers. The company was able to debug the design in the factory quickly, because IBM's fab engineers could work on one part while the Chartered engineers could debug a different part of the chip. They fed the information to each other, speeding the cycle of revisions. By Jan. 30, 2005, IBM tapped out the final version of the microprocessor.

ATI, meanwhile, had a more difficult time. The company had assigned 180 engineers to the project. Although games ran on the chip early, problems came up in the lab. Feldstein said that in one game, one frame of animation would freeze as every other frame went by. It took six weeks to uncover the bug and find a fix. Delays in debugging threatened to throw the beta-development-kit program off schedule. That meant thousands of game developers might not get the systems they needed on time. If that happened, the Xbox 360 might launch without enough games, a disaster in the making.

The pressure was intense. But Neil McCarthy, a Microsoft engineer in Mountain View, designed a modification of the metal layers of the graphics chip. By doing so, he enabled Microsoft to get working chips from the interim design. ATI's foundry, Taiwan Semiconductor Manufacturing Co., churned out enough chips to seed the developer systems. The beta kits went out in the spring of 2005.

Meanwhile, Microsoft's brass was worried that Sony would trump the Xbox 360 by coming out with more memory in the PlayStation 3. So in the spring of 2005, Microsoft made what would become a fateful decision. It decided to double the amount of memory in the box, from 256 Mbytes to 512 Mbytes of graphics Double Data Rate 3 (GDDR3) chips. The decision would cost Microsoft $900 million over five years, so the company had to pare back spending in other areas to stay on its profit targets.

Microsoft started tying up all the loose ends. It rehired Seagate Technology, which it had hired for the original Xbox, to make hard disk drives for the box, but this time Microsoft decided to have two SKUs — one with a hard drive, for the enthusiasts, and one without, for the budget-conscious. It brought aboard both Flextronics and Wistron, the current makers of the Xbox, as contract manufacturers. But it also laid plans to have Celestica build a third factory for building the Xbox 360.

Just as everyone started to worry about the schedule going off course, ATI spun out the final graphics chip design in mid-July 2005. Everyone breathed a sigh of relief, and they moved on to the tough work of ramping up manufacturing. There was enough time for both ATI and IBM to build a stockpile of chips for the launch, which was set for Nov. 22 in North America, Dec. 2 in Europe and Dec. 10 in Japan.

Flextronics debugged the assembly process first. Nick Baker traveled to China to debug the initial boxes as they came off the line. Although assembly was scheduled to start in August, it didn't get started until September. Because the machines were being built in southern China, they had to be shipped over a period of six weeks by boat to the regions. Each factory could build only as many as 120,000 machines a week, running at full tilt. The slow start, combined with the multiregion launch, created big risks for Microsoft.

An Unexpected Turn

The hardware team was on pins and needles. The most-complicated chips came in on time and were remarkable achievements. Typically, it took more than two years to do the initial designs of complicated chip projects, but both companies were actually manufacturing inside that time window.

Then something unexpected hit. Both Samsung and Infineon Technologies had committed to making the GDDR3 memory for Microsoft. But some of Infineon's chips fell short of the 700 MHz specified by Microsoft. Using such chips could have slowed games down noticeably. Microsoft's engineers decided to start sorting the chips, not using the subpar ones. Because GDDR3 700 MHz chips were just ramping up, there was no way to get more chips. Each system used eight chips. The shortage constrained the supply of Xbox 360s.

Microsoft blamed the resulting shortfall of Xbox 360s on a variety of component shortages. Some users complained of overheating systems. But overall, the company said, the launch was still a great achievement. In its first holiday season, Microsoft sold 1.5 million Xbox 360s, compared to 1.4 million original Xboxes in the holiday season of 2001. But the shortage continued past the holidays.

Leslie Leland, hardware evaluation director, says she felt "terrible" about the shortage and that Microsoft would strive to get a box into the hands of every consumer who wanted one. But Greg Gibson, system designer, says that Microsoft could have worse problems on its hands than a shortage. The IBM and ATI teams had outdone themselves.

The project was by far the most successful Nick Baker had ever worked on. One night, hoisting a beer and looking at a finished console, he said it felt good.

J Allard, the head of the Xbox platform business, praised the chip engineers such as Baker: "They were on the highest wire with the shortest net."

Get more information on Takahashi's book.

This story first appeared in the May issue of Electronic Businessmagazine.

Comments (0)

Please log in or register to post comments.
By submitting this form, you accept the Mollom privacy policy.
  • Oldest First
  • Newest First
Loading Comments...