AMD launched its first quad-core architecture, also known as Barcelona, back in September last year. Later on we also got a desktop version which we now know as Phenom. Much of the pre-launch information focused on the poor frequency scaling, which unfortunately turned out to be correct. Phenom is far behind its Intel counterparts in terms of frequency, and now that Intel is on par in terms of IPC (Instructions Per Cycle) AMD has to compete on other terms than pure performance.
AMD managed to beat Intel, and NetBurst, with the K8 architecture, but over the last year, the roles have been reversed and Intel is back on top. With another microarchitecture coming this fall, Intel is certainly enjoying being in the driver seat, but AMD intends to take over with the launch of the 45nm Shanghai core later this year.
AMD sources indicate that since the biggest updates of Nehalem are the integrated northbridge and the new interface QuickPath Interconnect, it has merely moved up Intel to where AMD has been for years. They did admit though that the improvement going from NetBurst to Core was an impressive one, although expected.
Even though we did not get to see any actual performance data, or even estimates, we were told that AMD is not especially scared by Nehalem. It considers HyperTransport to be equal or even better than QPI, while what it now has to do is find ways to make its processors run faster, not just higher frequencies but also get more work done per cycle.
Both are supposed to come when AMD starts shipping its 45nm processor Shanghai. Shanghai is the server processor and should go out to first and second tier partners in Q4. The new processor will sport 6MB of L3 cache which should prove beneficial in many scenarios. Previous tests with the Deneb core has at least pointed to this.
Unfortunately other sources indicate that Shanghai, and Deneb, will not hit significantly higher frequencies than Phenom. It should reach higher than Phenom did, but the roadmaps we’ve seen have only shown as high as 2.8GHz and that’s not enough if we don’t see more than what the Deneb preview showed us, which we very well may when server benchmarks are posted.
