With the fourth quarter now upon us, the market's gaze has begun to turn full force on AMD and the imminent launch of the company's Shanghai processor. Shanghai, for those of you who haven't been following AMD, is a 45nm die-shrink of the company's Barcelona architecture with a 6MB L3 cache (up from 2MB on Barcelona), improved performance and power characteristics, and an additional set of as-yet-unspecified performance tweaks. On Monday, Ars had a chance to sit down with Patrick Patla, head of AMD's workstation and server division, who offered insight into Shanghai's development process, as well as a few tantalizing details on what we can expect when the new processor ships.
According to Patla, Shanghai will offer 35 percent more performance and 35 percent better power efficiency when compared to Barcelona products. There appears, however, to have been a miscommunication at some point; CNet quotes Patla as saying: "At the same frequency (speed), Shanghai will outperform Barcelona by about 20 percent." When he spoke to Ars, Patla clearly and distinctly emphasized 35 percent improvements in both raw performance and power efficiency when comparing Shanghai to Barcelona. Regardless of which percentage is accurate, the point to take away here is that Shanghai will pack significantly more punch, clock-for-clock, than Barcelona ever managed.
If AMD actually delivers a consistent performance boost at either figure (20 percent or 35 percent) for any type common server workload, it will be a huge achievement. Evolutionary processor designs don't typically deliver this type of speed increase across the board; such unilateral gains are typically reserved for new simultaneous platform refreshes and/or CPU architectural debuts. When questioned on how Shanghai could deliver such improvements over Barcelona while remaining an evolutionary design, AMD emphasized a number of changes the company made back when Shanghai had yet to tape out. In the rush to deliver Barcelona, Patla implied, the company had actually gotten ahead of itself, and wrote finis to the project before it should have. When it came time to build Shanghai, AMD made a conscious decision to step back, take its time, and emphasize design over speed. This slower, more deliberate design process has resulted in a better processor, with Shanghai delivering at the top end of the target specifications AMD had set for it. This is not 2007's Barcelona.
AMD quad-core 45nm die
As for volume production, AMD is executing on that front as well. The company has consistently promised that Shanghai-based servers will ship before the end of the year, and an HP rep was on hand to confirm those plans. Neither company is offering guidance on CPU clockspeeds, but Patla did confirm that AMD will ship midrange 55W TDP processors first, followed by additional 105W parts in the first quarter of 2009. If these upcoming 55W parts actually are mid-range CPUs, it further implies that Shanghai's power consumption has improved dramatically. When Barcelona first launched last September, the Opteron 2347HE ran at 1.9GHz and carried a 68W TDP rating, the 2GHz Opteron 2350 was a 95W part, and the 2.5GHz Opteron 2360SE was rated at 120W.
Barcelona's TDP ratings have improved since last September; a modern 2360SE built on the B3 stepping now draws just 105W. Based on current information, we can expect TDP levels at any given clockspeed to drop significantly—Barcelona's 105W part may be a 90W or even 75W chip when built on 45nm. This in turn opens the door for faster Opterons, though AMD's comments imply that we won't see these high-end parts until after the turn of the year.
For all its strengths, those of you hoping Shanghai will vault Penryn entirely and drop squarely into Nehalem territory may well be disappointed. AMD told Ars this week that Shanghai is explicitly and directly targeted at Intel's quad-core, Penryn-based Harpertown processor. When queried on how the new chip will perform against Dunnington, AMD stated that relative performance between the two should be "good," but didn't indicate any miracles would be forthcoming. As for Nehalem, AMD emphasized the fact that Intel's next-generation processor will initially only be available in single and dual-socket configurations, while it intends to offer Shanghai in 1P, 2P, and 4P systems straight off the launchpad. Read between the lines, and the company is saying: "Don't look at our 1P or 2P performance, because we may well get slaughtered." AMD also has its own Dunnington competitor in the works, but don't look for the chip anytime soon; Sunnyvale's hexa-core processor won't arrive until Q4 2009.
Conclusion
Unless AMD is flat-out lying, it seems safe to conclude that Shanghai will deliver what Barcelona promised and then some. What's not so clear, unfortunately, is whether or not Shanghai alone can help AMD out of the tremendous hole the company has dug itself. Even assuming that all of the market issues work themselves out, AMD will still be playing catch-up to Intel in a number of market segments. Even with a top-flight Shanghai, AMD's margin for error is now effectively zero.
