Intel sticks to dual-die processors
By
Tom Sanders
24 October 2006 03:29PM
Tags:
intel | sticks | dualdie | processors
Two packages on a chip will help AMD to regain chip lead.
Intel plans to continue making its multi-core processors by combining dies on a single package beyond 2007, the company said at a meeting with reporters in San Francisco.
Intel currently manufactures its quad-core processors by combining two dual-core chip dies on a single chip.
AMD, by comparison, uses a monolithic design where all processor cores are manufactured on the same single die. This allows for memory sharing between the cores, making for faster processors that consume less power.
"When you design the product to be a more coherent quad-core system, then you get improved performance compared to putting two dual-core chips in the same package," Nathan Brookwood, an analyst with Insight 64 told vnunet.com.
Intel used the same dual-die approach when it first launched its dual-core processors codenamed Smithfield and Presler, but quickly created a monolithic die where both processor cores were manufactured together. This time it will not switch to a monolithic design that quickly, the company said.
"This [multichip approach] is not something that we just did temporarily. It's a design implementation choice that gives us real benefit in terms of time to market and volume and cost," said Stephen Smith, vice president for Desktop Platform Operations with Intel.
Manufacturing a monolithic chip is about 15 percent more expensive than dual-dies combined on a single chip, Smith argued, because of production yields.
Dual core processors have a smaller die size than a monolithic quad-core one would have. Because a percentage of all chips are damaged during the production process and have to be discarded, a large die is more likely to result in broken chips.
As chip production techniques advance and transistor sizes shrink, Intel expects to hit a crossover point where it becomes cost effective to produce single-die quad-core chips, but that can take years.
"I wouldn't be surprised if that is between 45 nanometer and the generation after that," said Smith.
Intel is scheduled to start producing 45 nanometer (nm) chips in 2007. The next generation of 32nm chips is expected to arrive around 2009.
According to data from Insight 64 analyst Brookwood, Intel's current dual-core design measures about 143 square millimeters. A monolithic quad core design would measure about 300sq mm. Brookwood by comparison expects AMD's forthcoming quad core chip to be about 200 sq mm.
Intel's chips are bigger because they have more cache memory on the die. The larger cache allows Intel to compensate for its chip's inability to share cache memory between the two separate dies.
The experience with the Smithfield and Presler demonstrates that the dual-die approach results in a performance drop, said Brookwood.
"When those chips had to coordinate, they had to go outside the chip. Now Intel is doing the same thing, but between the two dual cores. That eats up performance," Brookwood told vnunet.com.
Intel, however, is forced to take this approach because of its Front Side Bus design, Brookwood claimed.
"They have to work really hard to match AMD'S performance given that they still have the front side bus. That translates into bigger chips, bigger caches, and more complicated motherboards. Everything gets more complicated because of the front side bus."
But as AMD is not held back by these design constraints, the company will be able to create faster and less expensive chips. Brookwood projects that AMD will once again claim the lead in the performance per Watt metric when it releases its quad core chips next year.
"Even though Intel may be the same in the cost of its processor chips, every other aspect of Tigerton will be more complex and more expensive than that of the AMD platform."
Copyright © 2008 vnunet.com