Single page Print

A first look at Intel's quad-core Xeon processors

Clovertown comes to, er, nevermind

INTEL CHOSE TO DEBUT its first quad-core processor on the desktop with the Core 2 Extreme QX6700, but quad-core's most immediate potential lies with Xeon processors built for servers and workstations. There is no shortage of multithreaded code in the enterprise world, and plenty of applications—running either natively or on virtual machines—are ready to take advantage of additional processor cores. More is at stake in the enterprise world, too. There, higher performance does more than smooth out in-game frame rates and lower media encoding times; it can increase productivity, yield a competitive advantage, or reduce the time it takes to answer important scientific questions.

As it has on the desktop, power consumption has also become an increasingly important factor for servers and workstations. In the enterprise world, the cost of a processor is more than the price tag of an individual chip; it also takes into consideration the air conditioning needed to cool it, the electricity consumed by running it, and the real estate cost associated with the rack where it will live.

The Core microarchitecture is particularly power-efficient, and Intel's manufacturing prowess has quickly spun it into a quad-core design for the Xeon 5300 series. These new Xeon chips fit into the same power envelope as their dual-core predecessors, but pack twice the processor cores. Read on to see what makes the Xeon 5300 series tick and how new chipsets and motherboards will take advantage of its potential.

Clovertown in action: benchmark results soon

Introducing Clovertown
Intel's product codenames are pulled from the local geography of the Pacific Northwest, so they tend not to have much to do with the processors they represent. However, the "Clovertown" codename for quad-core Xeon processors is rather appropriate. Four-leaf clovers are the lucky kind, of course—although luck had little to do with this quad-core implementation.

As with its first dual-core desktop processors, Intel's initial quad-core Xeon designs put two dual-core dies on a single package. Opteron loyalists have been quick to note that this isn't a "true" quad-core implementation, and there are certainly some drawbacks to the design; dual-core dies must communicate with each other over an increasingly taxed front-side bus, for example. Still, this is the only quad-core design available on the market, and it's one that Intel has become quite adept at manufacturing.

Indeed, the company's manufacturing mojo is almost beyond reproach. Intel released its first 65-nano chips a year ago, and says yields are currently higher than they have been with any other process node. Yields are particularly important for quad-core designs, since doubling the number of cores increases the die area by a proportional amount. With a monolithic quad-core design, there's a much greater chance that a crippling defect will fall somewhere in the die area. However, by doubling up on dual-core chips, Intel can mix and match dies from anywhere on the wafer, or even from different wafers. Intel claims this approach improves yields by over 20% per wafer—much more than the "reasonably negligible" performance hit Intel says it takes going from a monolithic to dual-die design.

Quad-core Xeon 5300 series processors share the same 65nm manufacturing process and dual-die design as Intel's Core 2 Extreme QX6700 desktop chip, but there are several key differences. The Xeons come in an FC-LGA6 "Socket 771" package that isn't compatible with LGA775. They also support dual-processor configurations, making it possible to build an eight-core system using only two sockets. Xeon processors also have access to a faster 1,333MHz front-side bus than their desktop counterparts, and they're available in more configurations than Intel's desktop quad-core part.

Clock speed L2 cache Front-side bus TDP Price
Xeon E5310 1.60GHz 8MB 1066MHz 80W $455
Xeon E5320 1.86GHz 8MB 1066MHz 80W $690
Xeon E5345 2.33GHz 8MB 1333MHz 80W $851
Xeon E5355 2.66GHz 8MB 1333MHz 120W $1172

Fittingly, Intel is launching four quad-core Xeon 5300 models to start. Each chip will be equipped with a total of 8MB of L2 cache (4MB per die), and all but the 2.66GHz E5355 will fit within an 80W Thermal Design Power (TDP). Most of the E5300 lineup is optimized for what Intel calls efficient performance, or attractive performance-per-watt characteristics. The E5355, however, is aimed at performance-driven workstation users who will gladly swallow higher power requirements to get a faster chip. Think of it as an Extreme Edition Xeon without the requisite Mountain Dew commercial.

Although AMD and Intel have different methods for rating the TDP of their processors, it's worth noting that the bulk of AMD's dual-core Socket F Opteron 2000 series CPUs carry a 95W TDP. That line's flagship Opteron 2220SE 2.8GHz part has a TDP of 119W.

Only two of the Xeon 5300 series processors Intel is introducing today are equipped to take advantage of a 1,333MHz front-side bus. The rest are stuck at 1,066MHz, and it's those products that are ramping first. 1,333MHz parts are due to ramp early next year for OEMs, although we have a couple of E5355s in-house for a full review you should be able to read soon. The first quarter of next year will also see the introduction of new 1,333MHz models, in addition to a low-voltage quad-core Xeon that will run at 1.6GHz with just a 50W TDP.

A question of chipsets
Xeon 5300 processors should have no problems plugging into motherboards based on Intel's current Bensley platform. Bensley, of course, features dual front-side busses (one per CPU) at speeds of 1,066 or 1,333MHz, and supports fully buffered DIMMs (FB-DIMMs). We covered the Bensley platform when Intel rolled out the first "Woodcrest" dual-core Xeon processors, so it's nothing new. However, quad-core Clovertown CPUs will also work with Intel's upcoming Stoakley platform, which is due to debut in the first quarter of next year.

The Stoakley platform is based on a 90-nano shrink of the Bensley architecture. Bensley's dual 1,066/1,333MHz front side busses return, this time with support for upcoming 45-nano Penryn chips. Stoakley also features plenty of PCI Express, with 44 lanes of PCIe joined by a pair of second-generation PCIe x16 links. Generation one PCIe links can be used to hook into a variety of peripheral chips to provide Serial ATA RAID, Gigabit Ethernet, and PCI-X connectivity.

Seaburg is the codename for Stoakley's Memory Controller Hub (MCH), which features four channels of FB-DIMMs at 533 or 667MHz. Up to 128GB of memory is supported—double that of Bensley's Blackford MCH. Seaburg also offers an enhanced memory controller that Intel says improves sustained throughput by 25% and a larger, smarter snoop filter optimized for quad-core chips.

The snoop filter stores coherency information on all of the cache lines mapped to system memory. The Greencreek MCH's snoop filter isn't really optimized for quad-core designs; affinity groups are tied to each front-side bus, but quad-core CPUs add additional agents to those busses. With Seaburg, Intel has designed a snoop filter with four affinity groups—one for each last level cache on the front-side bus. Seaburg's snoop filter also features a new eviction algorithm, and its size has been increased to ensure complete coverage for future quad core-chips that Intel says will feature larger caches.

Technically, the Stoakley platform is only a part of Intel's workstation chipset roadmap. However, system builders are free to use the chipset in servers, if they wish.