New AMD server roadmap includes 4-, 6-, 12-core CPUs

— 5:17 PM on May 7, 2008

AMD has revised its server and workstation roadmap, and as CNet reports, the chipmaker's new game plan postpones the arrival of a next-generation processor architecture into the next decade. In 2009, rather than push its Bulldozer architecture as it originally planned, AMD now intends to stick to 45nm four- and six-core server CPUs.

The four-core CPU is code-named Shanghai, and we've already discussed it some. Shanghai will be based on a 45nm process and will pack 512KB of L2 cache per core, 6MB of shared L3 cache, DDR2 memory support, and some architectural tweaks. Istanbul, which will follow in the second half of 2009, will essentially be a remix of Shanghai with six cores instead of four. CNet doesn't mention whether Istanbul will be a single-die part or not, but previous rumors about the CPU suggest not.

In the first half of 2010, AMD plans to follow up with a revamped six-core chip dubbed Sao Paulo and a new 12-core offering nicknamed Magny-Cours. Both CPUs will be based on the same 45nm process as Shanghai and Istanbul, but they'll have support for DDR3 memory and include four-way HyperTransport 3.0 links. Where Istanbul and Shanghai will fit into existing Socket F servers, Sao Paulo and Magny-Cours will come hand-in-hand with a new Maranello platform, as well.

Interestingly, AMD's new roadmap is void of eight-core offerings. Intel, by contrast, has clearly stated plans to introduce eight-core Nehalem processors next year. Considering Nehalem could very well have higher clock-for-clock performance than Shanghai, AMD may opt to compete against four-core Nehalems with six-core Istanbuls and stay out of the eight-core game altogether.

Bulldozer hasn't quite vanished from AMD's plans. The company's executives stated last month that the upcoming architecture will sample next year on a 45nm process. Presumably, that means a launch isn't planned until the second half of 2010 at the earliest.

Tip: You can use the A/Z keys to walk threads.
View options

This discussion is now closed.