Original URL: http://www.theregister.co.uk/2009/06/08/amd_istanbul_window/

AMD Istanbul - Time for something new already?

A question of qualification

By Cade Metz and Timothy Pickett Morgan

Posted in Servers, 8th June 2009 07:11 GMT

Did anyone notice that AMD's roadmap puts maybe ten months between the company's 'Istanbul' Opteron and its next-gen server platform? That's an awfully small window when you're looking to qualify a multi-core server chip.

AMD has delivered the six-core Istanbul several months early, but the question is whether it will have a good ramp. In other words: Will it make some money? Technology isn't the issue. In this case, the issue is timing.

Launched last week, Istanbul plugs into existing Rev F machines as well as brand new models, such as HP's ProLiant G6 also launched last week. The six-core chip can stand toe-to-toe with Intel's 'Nehalem EP' Xeon 5500s for dual-socket servers, and chances are, it will best the 'Dunnington' Xeon 7400s as well.

That AMD has pulled Istanbul into initial shipments during May should mean it has a longer life than you might expect. But, as we reported back in April, the 8- or 12-core 'Magny-Cours' Opterons with their new G34 sockets are now scheduled for delivery the first quarter of 2010.

They started sampling this spring, and an official release is only 8 to 10 months away. Meanwhile, the 4- to 6-core 'Lisbon' Opterons with their C32 sockets will arrive in the second quarter of 2010.

Yes, after years of keeping the Opteron 2000 (two-socket) and Opteron 8000 (for four socket and larger) processors essentially the same, AMD is bifurcating its product line. Much like Intel has done with Xeon DP and MP platforms and now Xeon EP and EX platforms. Paired with those Magny-Cours processors (to be given the Opteron 6000 designation), the G34 platform is known as 'Maranello.' Joined by Lisbon (Opteron 4000), the C32 platform is dubbed 'San Marino.'

AMD has not given out the specs for the G34 or C32 sockets. But it's safe to assume that the G34 will have more than the Rev F socket's 1,207 pins and that the C32 will probably be a lot closer to the Rev F, maybe akin to the difference between an Opteron 1000 and an AM2 Athlon socket. (i.e. not a heck of a lot, but enough to keep separate platforms separated).

There's some talk from last year that the G34 socket has 1,974 pins. We're going to make the wild guess that the C32 will have 1,208 pins - one more for the extra HT3 link coming out of the package compared to the Rev F chips.

The point is: AMD is eager to blunt Intel's attack with four-socket 'Westmere' Xeon EP processors using 32 nanometer processes (due around the middle of 2010) and with eight-core 'Nehalem EX' Xeon 7500s (due to ship before the end of the year out of Intel's fabs and expected in systems in early 2010. But will anyone spring for Istanbul with Maranello and San Marino so close behind?

Qualifying a multi-core processor can take many months, and server customers typically standardize on a platform for a year to a year and a half. If you're going to spend the time, effort, and money to qualify new hardware, why not wait for Magny-Cours?

"In Intel's experience, a shorter product lifecycle can make it harder to ramp that product," Intel told us after we asks for comment on its rival's tiny Istanbul window.

"Customers tend to enjoy stability as well as performance. We have obtained direct feedback from our customers that getting single-socket systems to market takes a couple months of qualification, two-socket servers take a few months, and multi-socket servers can take up to six months to reach production qualification."

Qualification is in the eye of the beholder

Naturally, AMD says it's unwise to write off Istanbul. John Fruehe, director of business development for server and workstation products at AMD, argues that because it plugs into an existing platform, Istanbul will requires a relatively short qualification. Maybe three months. Meanwhile, Maranello will take significantly longer to qualify - thus widening the gap beyond 8 to 10 months.

"Because customers are going to take a while to qualify these Magny-Cours and Lisbon platforms, you'll expect to see Istanbul having a pretty good lifecycle," Fruehe tells The Reg. "Last quarter, we did a fairly good business still selling some dual-core processors, because people standardize things and they don't like to go ripping things out and changing things."

That's one way to spin it. But the new platform isn't that different from the existing setup. The guts of the Magny-Cours and Lisbon processors are very similar to the company's current chips - the cores, caches, and so forth - even if the memory controller is being upgraded to DDR3 support and the sockets are being differentiated from the current Rev F.

Whatever the case, Fruehe says Istanbul will enjoy a healthy life after the debut of the Magny-Cours family. "We expect there will still be a market for Istanbul, whether it's somebody who has already qualified it and that's what they want to keep deploying, or someone who is in the process of deploying the new platform but just isn't ready to make that move yet."

The Rev F socket was introduced in August 2006, and it has a five-year lifespan, so that means AMD will keep supporting it until the summer of 2011 - and possibly longer if the transition to the G34 and C32 platforms takes a while. But that probably does not mean the company will create an eight-core kicker to Istanbul.

AMD never had any plans of creating an eight-core Rev F processor. Even the eight-core 'Montreal' Opteron - supposedly killed off in March 2008, when AMD decided to create two distinct Opteron families with distinct chipsets and core counts - was intended to be a Socket G chip.

Whether the sales opportunity window is small or not, AMD has little choice to try to shake down revenues - even in this poor economy. Just like Intel did with its less-than-impressive 'Dempsey' Xeon chips, which debuted in May 2006 with an old processor architecture crammed into a new server platform, a platform that didn't really make sense until the kicker 'Woodcrest' Xeons came out in June of that year with much better performance and thermals. And the Woodcrest chips didn't start appearing in platforms until September.

"So the pace we're going at is not that different from our competitor," says Fruehe, "and because Istanbul is in the same Rev F socket, we think it's less disruptive to developers."

AMD will sell whatever it's got on the truck - just like Intel has done in the past and will do in the future.

And maybe the processors aren't the thing to worry about. With neither Broadcom nor nVidia making chipsets that compete with the G34 and C32 - and AMD coming out with its first chipsets since the original Opteron launch back in 2003 - the chipset support for the Magny-Cours processors could be more of a concern than anything else.

If AMD gets the chips right but flubs the chipsets, that will be a disaster for the company and a cause for celebration for Intel. Keep an eye on that 'Fiorano' platform and AMD's homegrown SR5690/SP5100 chipset. They're designed specifically for the Istanbul processors, and they're undoubtedly a dry run for the G34 and C32 chipsets. ®