|
||
|
|
|
              |
|
 |
 |
||||||||||||
 |
|
|||||||||||
By Chris Angelini |
||||||||||||
If I were to tell you that Intel's server and workstation chips have been matched evenly against AMD's Opteron processor, I'd be lying. Sure, the company had a strong Xeon brand back in 2003 when the K8 architecture first launched. But even after AMD's drive to promote 64-bit computing slowed in the face of delayed mainstream OS support, it was able to fall back comfortably on superb 32-bit performance, an integrated memory controller, and a super-efficient HyperTransport interconnect. The story only got better as time passed. When both processor vendors announced their respective plans for dual-core, AMD was the one to say its Opteron had been conceptualized with multi-core processing in mind. For the most part, Intel kept a low profile with its first dual-core server chips, apparently in anticipation of something bigger and better. Now, that something is ready for battle with the best AMD has to offer. But it isn't Intel's style just to announce a new processor design. Instead, we're seeing an entirely fresh platform, spanning CPUs, chipsets, and motherboards, all developed with one objective in mind—to put Intel back on top of the server and workstation game. Dempsey Emerges
Intel's first dual-core Xeon processor was forged using NetBurst steel, meaning it borrowed the Pentium D's architecture—the one Intel has massaged since late 2000. Notoriously power-hungry and hampered by the bandwidth constraints of a single 800 MHz front side bus, the Paxville core was never really positioned as Intel's dual-core opus. It was more of an appetizer for customers who demanded Intel hardware and stood to benefit from multi-core technology. The entrée was to be Dempsey, now known more officially as the 5000-series Xeon processor. Manufactured on a relatively fresh 65 nm process and consisting of two distinct cores on a single package, Dempsey shares a lot of technical similarities with the desktop Presler design, manifest in the Pentium D 900-series lineup. It's still a steed from the NetBurst stable, but by shifting to 65nm and cranking up the clock frequency, Intel is able to improve Dempsey's performance per watt versus the preceding single-core Irwindale offering. The two cores on every Xeon 5000-series processor each sport 2MB of cache for a total of 4MB per package, just like Presler. They also support Intel's 64-bit extensions and the Execute Disable bit—again, just like Presler. The new Xeon sets itself apart, however, through a 1,066 MHz front side bus and the addition of Hyper-Threading. "But aren't those features touted by Intel's higher-end, dual-core Extreme Edition processors?" you ask. Indeed, they are. And in fact, the flagship Xeon 5080 runs at 3.73 GHz, similar to Intel's fastest Pentium Extreme Edition chip. The new Xeon does come equipped with a couple of unique tricks to keep it differentiated, though. Most obvious is dual-processor support. Pentium D and Extreme Edition chips only support single-socket LGA775 configurations. Although an Extreme Edition includes two cores and Hyper-Threading, causing Windows to recognize a total of four processors, there's simply no way to plug two of those CPUs into a motherboard. The Xeon 5000-series, on the other hand, leverages an all-new LGA771 interface that works with dual-processor boards. Thus, outfitting a server with two 5000-series Xeons gives your customer the ability to churn through eight threads simultaneously. Virtualization also emerges as a key selling point, since neither the single-core Irwindale nor dual-core Paxville designs support it. Not even AMD's Opteron includes hardware virtualization at this point, though the technology will be enabled in the months to come. Until then, Intel is encouraging resellers to explore the potential of virtualization, which facilitates multiple operating systems and applications on a single hardware platform. That's a particularly valuable proposition for businesses able to benefit from cost-cutting server consolidation. Waiting for Woodcrest If Dempsey is the entrée keeping you and your customers satiated with plenty of performance, then the upcoming Woodcrest is dessert—a cherries jubilee on which nobody will want to pass. Intel claims that Woodcrest will begin shipping in June under the Xeon 5100-series moniker, ushering in an even sweeter balance between performance and power.
The boost comes from a long-awaited shift toward the Core microarchitecture and away from NetBurst. Whereas the 5000-series chips maintain a 130W maximum TDP, 5100-series Xeon processors max out at 80W. And Intel claims that its 5100-series Xeon will perform up to 125% faster than the previous generation of dual-core Xeons. If you haven't yet heard of Conroe and Merom, those are the desktop and mobile Core-based processor designs expected to catapult Intel to the top of the performance charts later this year. Woodcrest is cast from the same die, so your customers should expect similarly compelling speed. Like Dempsey, Woodcrest is manufactured on Intel's 65 nm process. Rather than feature two separate pieces of silicon sitting side by side on a package, the chip is singular, incorporating 4MB of cache shared between two execution cores. Preceding dual-core Xeons were slowed somewhat by constant cache updates—one processor letting the other know what it was storing in cache—traveling over the front side bus. Woodcrest's shared cache helps mitigate some of that traffic. Intel's Core architecture takes a step back from the emphasis on clock speed, opting instead to focus on efficiency. Thus, you'll see frequencies up to 3 GHz initially, with models scaling all the way down to 1.6 GHz. The Power of a New Chipset The Dempsey and Woodcrest processor designs are instrumental components in Intel's scheme to regain SMB leadership. Had the two cores been constrained to older chipsets, however, there's a good chance that neither would be able to usurp Opteron's pole position. The problem is one of bandwidth. Whether you're running a single processor or eight, any call to system memory has to travel over a front side bus, through the northbridge, out to RAM, and back again. Continue adding processors to the bus and it's easy to see how that pathway might become congested. AMD gets around the issue by using HyperTransport interconnects and a built-in memory controller to create a mesh of sorts, where each CPU is linked to the others. The obvious downside is that any change in memory technology support requires a major processor revision. That's why Opteron's memory controller hasn't seen a major overhaul since it emerged in 2003. Intel continues to favor the path of flexibility by maintaining a separate memory controller on its new 5000P chipset, formerly referred to as Blackford. Of course, that leaves an escalating bandwidth issue as you scale from a single-core processor to a dual-core chip to a pair of dual-core CPUs and so on. Consequently, the Bensley platform calls for dual independent front-side buses in an attempt to augment throughput. Each bus runs at 1,066 MHz or 1,333 MHz, depending on the processor installed, yielding up to 10.5 GBps per bus.
The 5000-series chipsets also showcase FB-DIMM (Fully Buffered DIMM) memory technology for the first time, transitioning yet another sub-system from parallel to serialized data transfer a la SATA storage and PCI Express. In essence, FB-DIMMs utilize faster and narrower data paths. Cutting back on the number of electrical traces helps free up space on complex motherboard designs, which can then be used to add even more memory slots. You can consequently offer customers high-end configurations with up to 64GB of memory, up from the previous 16GB limit. An FB-DIMM module actually consists of standard DDR2 memory behind an AMB (advanced memory buffer) chip, which enables serial communication between the module and memory controller. Added complexity will undoubtedly make FB-DIMMs more expensive individually. But as you approach higher memory capacities, extra memory slots will make capacity upgrades less expensive since you won't be forced to buy ultra-pricey 2GB and 4GB modules. In the long run, Intel sees FB-DIMM technology as a mechanism for reducing costs. Finally, the 5000-series chipsets feature three x8 PCI Express links, for peripheral devices, such as Intel's 6321ESB I/O controller or the 6700PXH PCI hub. Bensley in the Channel
Intel finalized the hardware behind its Bensley platform a while ago, as we've been hearing about Dempsey processors up and running in labs for months. In the time before launch, Intel has busied itself by laying the foundation for a successful channel debut. Product is consequently shipping and available, Intel claims. There are two SKUs in particular that Intel says work well as channel plays, both priced below existing single-core Xeon processors. The first is a 2.67 GHz boxed model (Xeon 5030) available exclusively to resellers. It sells for less than $180 and offers a conservative 95W TDP, giving VARs an inexpensive avenue into multi-core, multi-processor servers. The Xeon 5050, running at 3 GHz, is another favorite. Intel sells that one in a box or on a tray, giving it the same attractive 95W power rating and sub-$200 price tag. When an SMB customer shrugs off the idea of a server because of perceived cost, these two chips are your aces. A dual-socket server with just one Xeon 5000-series chip brings plenty of performance to bear, and the spare processor interface is a cheap insurance policy should more speed be needed down the road. Intel is also shoring up FB-DIMM availability for its channel partners. Five memory vendors are currently supplying modules to the channel with immediate availability. Unfortunately, a combination of low volume and the AMB chip are resulting in premium pricing. But Intel spokeswoman Mary Ragland says the company is working on a rebate program, available through distribution, to help resellers get their hands on the newer memory technology more affordably. Of course, Intel is offering its own lineup of motherboards, chassis, and barebones kits, which the company claims should already be shipping into the channel. By enabling the entire Bensley infrastructure, from CPUs to memory to enclosures, VARs get immediate access to all of the hardware necessary to turn an exciting announcement into sellable product. There aren't many vendors capable of pulling off such a feat, but Intel's efforts really serve resellers striving to combine performance and value at the SMB level.
Additionally, ASUS, Supermicro, and Tyan have all unveiled their own motherboards and chassis systems, addressing compact 1U form factors, larger 5U pedestals, and everything in between. Back on Track We haven't yet had the chance to put Bensley's performance to the test, so drawing conclusions against the Opteron would be premature. However, Intel is in an unquestionably better position to compete for server and workstation market share now than it was a month ago. If all goes as planned, we'll be saying the same thing again a month from now as the much-anticipated Woodcrest-based Xeon pulls up alongside Dempsey. |
||||||||||||
 Back to top |
||||||||||||
Copyright © 2007 RAM Magazine. All rights reserved.
Do not duplicate or redistribute in any form. |
||||||||||||
