Intel's Nehalem takes microarchitecture rivalry to 45nm

At first glance, Intel Corp. appeared to have stolen a page from Advanced Micro Devices Inc.'s playbook when it revealed that its next-generation 45nm microarchitecture, due in production next year, will incorporate both integrated memory controllers and an optional integrated graphics engine. Analysts differed about the degree to which Intel's moves could be a direct reaction to AMD's, but the prevailing sentiment is that the result should keep the vendors in a tight competition, from an architectural point of view.

The integrated memory controllers and graphics engine are in the mix for Intel's forthcoming architecture, code-named Nehalem, said Pat Gelsinger, senior VP and general manager of Intel's digital enterprise group. Gelsinger provided few details, but did say Nehalem supports multithreading of up to 16 or more threads on eight or more cores with scalable cache sizes. Intel's current Core architecture supports up to four cores and does not support integrated memory controllers or graphics.

AMD did it first
AMD has been offering an integrated memory controller since the introduction of its Athlon 64 in 2003. Buoyed by its acquisition of graphics chip vendor ATI Technologies Inc., AMD said last year it plans to deliver X86 processors with on-board graphics in late 2008.

Some people are saying that Intel is following in AMD's footsteps with its latest plans for Nehalem, acknowledged Dean McCarron, principal of Mercury Research. But "Intel's argument for why it's doing this now as opposed to earlier does have some credibility," he said.

According to McCarron, Intel had a limited transistor budget and decided to allocate transistors to a larger cache memory rather than integrate the memory controller with the CPU. With Nehalem, designed for a 45nm process with high-k dielectrics that Intel says will double the transistor budget of its current 65nm process, "they now have the transistors to do both," he said.

Both companies are moving to integrated memory for the same reasons. "There is a pretty immediate performance benefit when you move to an integrated memory controller," McCarron said. "You are getting rid of a lot of latency."

Chris Casso, an analyst with Friedman, Billings, Ramsey and Co. Inc., said in a research note that Intel is "essentially copying the innovation that AMD pioneered about two years ago, which allowed AMD to gain share from Intel in servers." AMD had a great idea, but with Intel following suit, it will "erase AMD's advantage," Casso wrote.

Asked whether Intel was responding to AMD, Gelsinger said the general trend is to integrate more functionality with the microprocessor and that other companies had "probably arrived at the same conclusions" independently.

McCarron said the graphics integration would save power and space, particularly important for mobile computing. He said that the integration does not mean Intel will eliminate the external graphics attachment port.

Flopped efforts
McCarron noted that this is not the first time Intel is integrating the memory controller and graphics with a CPU, but "likely the first time it will do it successfully." Intel's Timna chip, which was cancelled in 2000 before coming to market, called for such integration. McCarron said Timna was axed because its Rambus memory interface did not come down to meet the requirements of the low-end applications it was designed to target.

On the day it completed the $5.4 billion acquisition of ATI last October, AMD detailed plans to offer its next-generation X86 microarchitecture, Fusion, in 65nm by late next year. It appears Intel plans to beat AMD to the punch with its next-generation X86 microarchitecture—and do it at the next process technology node.

"It's safe to say that the really direct competition between the two companies in terms of being really nose-to-nose shows all the signs of continuing in the future," McCarron said.

- Dylan McGrath
EE Times