Barcelona, slated for release later this year, is the next version of the Opteron and AMD’s answer to the dual-core Woodcrest and quad-core Clovertown Xeon processors introduced by Intel last year. Billed as a “true” quad-core chip, Barcelona has four processor cores on a single piece of silicon, or die, compared to Clovertown, which straps together twin dual-core dies inside a single package, yielding a quad-core processor.
While the difference may be meaningless to most users, the dual-die design allows Intel to get chips to market faster and eliminates the need for a more-complicated chip design, said Kirk Skaugen, the general manager of Intel’s Server Platforms Group, during a recent interview.
“I can make lots of arguments for why (you might want) more elegance (in the design) and higher performance, but if I’m nine months late you blow that whole advantage away,” Skaugen said, noting that a Barcelona die is likely more complicated and more expensive to produce than the dual-core dies used in Clovertown.
“I think Barcelona could be an intellectually elegant chip. Will the market care if its cost is higher and it can’t ramp into volume at the (clock) frequency targets? Well, we’ll have to see,” he said.
AMD, meanwhile, is betting that the single-die Barcelona design will yield lower power consumption and offer higher performance than anything Intel can bring to market this year.
“Originally, we had a big performance gap with Opteron. Woodcrest certainly narrowed that gap, but Barcelona will return us to a position where we have that big a gap again, if not bigger,” Kevin Knox, the vice president of AMD’s commercial business, said in January.
The stakes are high for Barcelona and AMD, which has seen Intel recently claw back lost ground in the server market.
In recent years, Intel has usually led industry transitions to more advanced manufacturing processes. More importantly, Intel engineers have a knack for getting good yields with each transition, a reference to the high percentage of usable chips that come out of its factories. High yield rates mean production lines are more efficient, which keeps costs low and insures product availability.
These strengths helped Intel regain lost market share with the Clovertown and Woodcrest chips.
Each successive advance in technology allows chip makers to produce processors that run faster and consume less power. They can also reduce the size of the chip, which means that more can be produced on a single silicon wafer, lowering unit manufacturing costs and boosting profit margins. Or they can choose to add more features, such as an expanded memory cache.
Most of Intel’s processors are produced today using a 65-nanometer process, which is one generation ahead of the 90-nm process used to make AMD’s chips. These numbers, measured in billionths of a meter, refer to the average feature size that can be created on a chip and smaller is better.
AMD is gearing up for a transition to a 65-nm process later this year, but Intel isn’t standing still. The company is ready to begin producing chips with a more advanced 45-nm process that will again extend its manufacturing lead over AMD. And next year Intel will revamp its processor line with a more advanced microarchitecture, called Nehalem.
This is Intel’s “tick-tock” strategy, which follows the introduction of a more-advanced process technology one year with a new microarchitecture the next. With Nehalem set for 2008, Intel’s roadmap shows plans to begin using a 32-nm process in 2009 and the introduction of another microarchitecure, called Gesher, in 2010.
Staggering the introduction of a new manufacturing process from the rollout of a new chip design in this way will allow Intel to produce new chip designs using a mature production technology. “By decoupling that, you reduce the complexity by 50 percent,” Skaugen said.
With Barcelona, AMD is taking the opposite approach, looking to bring out a new chip design on a new manufacturing process — a transition that any chip maker can struggle with, no matter how strong their manufacturing teams are, Skaugen said.
“I’m not saying it can’t be done but, man, it has risk,” he said.