Practical Chip Design

At the DAC design management track today, technical managers from three industry leaders—Intel, ST Microelectronics, and Media Tek—described the process of managing an advanced-node, low-power chip design. The result was three complementary perspectives on the challenge of design management: managing people, managing technology, and managing resources.

The lead-off paper, presented by Intel Mobility Group vice president Elenora Yoeli, gave an executive's-eye view of the management of the Intel Atom CPU development process. Starting with the results, Yoeli descibed Atom—perhaps more familiar to this audience as Silverthorne (if you're having trouble keeping Intel's names straight, try Brian Dipert's excellent analysis)--as a Core instruction-set-compatible CPU designed from a clean sheet of paper for low power. The CPU runs at 1 to 2 GHz, and one tenth the power of the next lowest-power processor in the Intel IA line-up.

Yoeli said that power savings began at the level of architectural planning—looking at all architectural alternatives, and rejecting anything that didn't yield at least 1 percent improvement in performance for 1 percent increase in power. For example, she said, out-of-order execution was rejected, because the incremental performance gain over in-order execution couldn't justify the substantial increase in power consumption. But hardware-supported multi-threading was included, offering a 40 to 50 percent increase in performance for a 10 percent increase in power.

Once the team had quantitative goals, Yoeli said, the next challenge was to manage a leading-edge design to minimize risk, stay on schedule, and meet those goals. She described a number of quantitative management tools that identified project status, the current probability of achieving the project goals and schedule, and the critical paths. "A very important thing was to manage the critical path," Yoeli said. "That means knowing which team is on the critical path at any given time, and protecting them from the larger Intel."

Yoeli cited two techniques that particularly helped in project management. One was extreme agility of resources. Teams were flexibly organized, with an individual's responsibilities going both vertically—for a portion of the design—and horizontally—for a particular skill set or portion of the flow. Resources had to be flexible enough to shift quickly. And, in the interest of both agility and control, Yoeli broke large tasks into very small tasks. For instance the team took what would ordinarily have been nearly two years worth of RTL design, broke it into small RTL modules with short cycle times and concrete milestones, and managed the whole set of small tasks to completion in one year. "One of our sayings was 'code it once,'" Yoeli said. "Don't write RTL for a block and then keep going back and adding functions. Make the block small enough to understand and then write it all at one time."

Another key point Yoeli made was that decision-making requires managing just as much as any other part of the project. "We defined our decision processes early. Then we were willing to spend engineering resources to get data, so we could make data-driven decisions. Then we let the technical people decide—not the managers," she said.

Finally, Yoeli pointed to the advantage of working with a small design team. The Silverthorne team was tiny by Intel standards, with just one layer of management. This vastly simplified controllability, observability, and flexibility for the entire process. And it helped in preparing the team to react to surprises. "This was an entirely new design. And we were one of the first two teams using the new 45 nm process," Yoeli said. "We had to plan for surprises."