INTEL SERVER TECHNOLOGY:
TOTAL REFRESH

The server market's growth remains strong overall, but one of the hottest bands within the server spectrum is the "value server," the affordable yet high-performance design best suited to SMBs. There are many hardware options available for resellers targeting this segment. We believe the best value proposition (never mind the highest availability) rests with Intel's newest wave of Xeon processors, chipsets, and motherboards. Whether you're targeting single pedestals or rack clusters, the innovations here will help you win business and boost your clients' growth.

History doesn't always repeat itself. In 2002, a combination of factors led to a setback in the fortunes of Intel's Xeon server/workstation processor line. With the Core microarchitecture revamp in early 2006 came a steep improvement in Intel's server efforts. The server/workstation space continues to be very competitive, but Andy Grove's famous quotation is truer now than ever. "Success breeds complacency. Complacency breeds failure. Only the paranoid survive." Since 2005, Intel has resumed a firm grip on its paranoid roots, and nowhere is this more apparent than in the top-to-bottom refresh of Intel's server product family.

PREPARE FOR PENRYN. Intel CEO Paul Otellini shows off a wafer of 45nm "Penryn" cores, such as those found in the latest Xeon processors. Penryn contributes to many of Xeon's key advantages.

You get a sense of the impact Intel's new lineup is having from just a quick glance at IDC's August 2007 Worldwide Quarterly Server Tracker. At a time when some people expected server sales to decrease because of supposed anticipation for competing, delayed products, server market revenue actually gained 6.3% year-over-year in 2Q07. The largest growth area within this server market was "volume" servers, which showed 11% growth and dwarfed the mid-range and enterprise segments. Buyers weren't waiting for anything. They were jumping on the many benefits of the dual- and quad-core architectures already shipping in mass quantities in Intel's Xeon lineup.

When the Bensley platform arrived in 2006, it was a pretty simple story: three slightly different chipsets, a few Xeon 5000-series SKUs, and little more. Today, the Xeon family tree has grown from a sapling into a many-branched oak. This is good news in that you have a full complement of products ready to cover any need from a humble 1P, single-task box to high-end, MP systems ready for large-scale clustering. The entire family is now based on a single architecture, which helps streamline your sales and support efforts. And know that pricing has reached very aggressive levels from top to bottom.

We're not going to pitch you on the wisdom of servers here. Few small businesses and certainly no mid-sized or enterprise companies would hope to exist in today's market without servers assisting their operations. The trick is to understand the server processor options available, how these options have been improved recently, and some of the strategy points in recommending them.

Starting Out: Xeon 3000
Roughly 40% of the whitebox servers sold are IP (uniprocessor) machines. Adoption in small businesses and microbusinesses (fewer than 20 employees) is climbing rapidly. The trouble is that many of these entry-level servers aren't really servers. They're desktops cloaked as servers.

Hop around the various tier-one sites to get a look at the present field. Here's one for $1,199 running a Pentium D processor, 1GB of 667 MHz DDR2 memory, an 80GB SATA drive, and Microsoft's SBS 2003 R2. The specs may be trailing edge, but the system sounds downright deluxe compared to this other unit: $559 for a Celeron D, 512MB, 80GB hard drive, and no software. You get the idea. These are servers only in name. You're throwing a single processor core with minimal memory at a low-demand task.

Even then, if that were the whole story, the the "desktop as server" idea might still be sound. After all, a desktop machine can function just fine as a low-volume mail server or Web server in non-critical, occasional-use situations. But businesses grow. Needs evolve. What if the traffic volume on that bargain basement mail server doubles or triples? What if a cheap nearline storage server crashes at the moment that matters most because the system had a 1% failure rate instead of a 0.01% failure rate? To paraphrase Wayne Gretzky, you have to skate to where the puck is going, not where it has been. Companies need to buy their computing assets with anticipated growth and reliability in mind.

These are two of the chief reasons a Xeon 3000-based system is the go-to choice for 1P servers. (Any code name followers out there may want to know that the third quarter's Xeon 3000/3200 lineup was part of the Kaylo platform. In the fourth quarter, this transitioned into the 3100/3300 Garlow platform.) You want the performance available in the Core architecture, which is something you won't get in old model Pentium D and Celeron products. By the same token, you want at least a dual-core processor. Single-threaded, single-application servers may still be common in companies whose operations revolve around a single task, such as a search engine provider. However, most small businesses buying volume servers benefit over the long term primarily from general-purpose systems able to accommodate changing tasks. Moreover, the old school model with 1P systems was to buy one box for one task. With dual- and multi-core processors combined with virtualization technologies, owners can consolidate multiple servers into one physical system with very little performance impact and no loss in stability.

SIMILAR BUT BETTER. The Xeon 3200 line is nearly identical to its Core 2 Quad counterparts. However, the support and ISV validation behind Xeon make it the obvious business choice.

The other element to consider is support. With a server/workstation processor, software vendors are more likely to validate the hardware for their products and offer immediate support when problems arise. For example, UGS Solid Edge lists the Intel Pentium in its "minimum" hardware requirements but shows the Xeon processor in its "recommended" list. You can hear how the support call would go. Support: "Sir, are you running an Intel Xeon processor?" Customer: "No, it's a Core 2 Duo." Support: "Ah. I'm sorry, sir, but we haven't validated our application for that hardware." Customer: "But it's the exact same chip as the Xeon." Support: "Perhaps. But I can't offer support on a product we haven't validated." And so on.

True enough, the Xeon 3000 series is a carbon copy of the Core 2 series, right down to the LGA775 socket and ability to only function in single-processor environments. They're the same chip with different labels. Fortunately, they're also the exact same price. So one of the main reasons to buy a Xeon 3000-based system is that there is potentially no price delta between a server and desktop system. Mind you, it makes absolute sense to upsell other components in the system to "server-class" status to complement that server CPU. ECC memory may make more sense than standard DDR2. An enterprise SATA drive may serve a business' long-term needs better than a regular SATA drive. When you sell a desktop box, you're competing with every mass merchant in the country. With a server, you can enter the sales conversation at the same pricing level, then migrate upward.

BUILT TO BE REAL. The S3000AH may look like a desktop board, but you can tell from features such as slot orientation, integrated graphics, and more than this board is built for commercial duty.

As with the desktop Core 2 family, there are two main sequences in the 3000 series: the dual-core 3000 group and the quad-core 3200 group. Within the 3000 series, there are now two strains: the original models with a 1066 MHz front side bus (identifiable by model numbers ending in 0), and the newer SKUs with a 1333 MHz bus (with models ending in 5). As of this writing in November, the majority of shipping Xeons are based on 65nm technology. The starting model is the 3040, decked out with a shared 2MB L2 cache, and a 1.86 GHz clock speed. At the 3060, you jump into 4MB of L2 cache complementing a 2.40 GHz clock. With the Xeon 3065, not only do you get that frequency increase on the front side bus, but you also keep the 4MB L2. This crossover area may seem a little tricky, but it's really not. The 3065 at 2.33 GHz offers a negligibly slower clock speed than the 3060 at 2.40 GHz, but in most situations, the 30% boosted FSB is likely to give the 3065 a performance edge over its slightly higher-clocked adjacent sibling, plus the 3065 is priced lower. The current dual-core 3000 line tops out with the 3085, sporting a 3.00 GHz core frequency, 4MB L2, and a 1333 MHz FSB. The entire 3000 series, including the latest 1333 MHz FSB parts, boast an energy-friendly 65W TDP rating.

Moving into the quad-core 3200 series, we start with the 2.13 GHz X3210 followed by the 2.40 GHz X3220 and the 2.66 GHz X3230. All current Xeon 3200 series models feature a 1066 MHz FSB and 8MB of L2 cache. Interestingly, the X3230, the fastest part of the three, specs out with the lowest power consumption–only 95W TDP versus 120W for the other two SKUs. Additionally, the X3230 is the only Xeon 3000 member to offer SSE3 support. This could prove very handy for accelerating servers or workstations dedicated to handling multi-threaded multimedia applications.

Xeon 3000 Chipsets
A CPU without a chipset is just a flashy key fob. The chipset is the central nervous system that enables the communication between nearly all of the components in a PC, workstation, or server. As such, you can't upsell suitable server-class components to match your server CPU unless you also have a server chipset in play to maximize each component's potential.

The Xeon 3000/3200 processor line is backed by three corresponding northbridge parts, all of which are designed to accompany the Intel ICH7 series of southbridges. The first of these is the 975X northbridge–the exact same chipset found in Intel's BadAxe line of enthusiast consumer motherboards. In fact, there's a copy of the BadAxe 2 found in Intel's workstation motherboard collection–the S975XBX2–right down to the three PCIe x16 graphics slots. This borrowing of the top-end consumer chipset as the starting point of the workstation/server northbridge family has been an Intel tradition for a few years.

DOUBLE THE POWER. Sometimes, having two discrete servers is better than one virtualized box. The X38ML platform offers two 1P server boards in a single 1U rack design.

The S975XBX2 dropped from Intel's roadmap during Q3, though, leaving the 3210 chipset-based S3200SH (see below) as the only Intel house-brand motherboard for 1P, general-purpose pedestal servers. This perhaps underscores the popularity of 1U server designs in today's SMB market.

The 975X played mostly as a workstation chipset and was something of a standout from the rest of the 3000 generation. The 3000 and nearly identical 3010 chipsets were the most common parts in Intel's 1P server ecosystem. The specs on these chipsets topped out at a 1066 MHz front side bus tied to an LGA775 socket and up to 8GB of 667 MHz DDR2 (ECC supported). In general, the 3000/3010 was very similar to the 975X save in one important way: The 3000/3010 was designed for more flexibility in slot deployment, including with PCI-X. You might also consider that boards such as the S3210SH are validated for server operating systems while boards based on the S975X workstation chip are not. We mention these parts here not to confuse you with more SKUs than necessary but because there is still plenty of product in the channel, and you should be aware of its attributes and limitations compared to newer chips before making a hardware recommendation.

Going For Garlow
October was a major transition point for the Xeon 3000 family. The 1333 MHz bus CPUs mentioned above started sampling in August, but there was no core logic to support the new FSB frequency until the recent arrival of the 3210 (Garlow) and X38 (Garlow WS) chipsets, both of which typically pair with the ICH9R southbridge. The 3210 isn't a major overhaul like we saw with the 5000 series (Bensley) launch. Rather, the 3210 is a smart, necessary iteration that will prepare your 1P servers to tackle the advances coming in 2008.

The speed jump to a 1333 MHz front side bus tops the 3210's list of killer specs, guaranteeing users can get the full potential of the latest Xeon SKUs. Alongside this, the 3210 introduces support for Intel's first wave of 45nm processors. In the 1P server world, these are the new dual-core 3100 (Wolfdale) and quad-core 3300 (Yorkfield) chips. The 3065 will remain as the base-level, dual-core 1333 MHz FSB part going into Q1 of next year, but the new, 45nm E3110 (3.0 GHz, 6MB L2, 65W) slips into position above it.

You'll see dual-core continue as the foundation of the 3000 family throughout 2008, but expect Intel to apply more effort in pushing quad-core into the mainstream. The X3210/20/30 processors stay current throughout Q4. In Q1, though, look for 45nm manufacturing to become the new norm. The quad-core X3320, X3350, and X3360–all of which spec out at 95W TDP–will swing into rotation and remain there throughout the year. The X3320 is a 2.50 GHz part with 6MB of shared L2 cache. The X3350 nudges up to 2.66 GHz but doubles the L2 cache to 12MB. Finally, the X3360 keeps the 12MB cache and hits 2.83 GHz.

After a 1333 MHz FSB and 45nm compatibility, the 3210 keeps the status quo by maintaining use of DDR2 (including ECC) on the 1P platform rather than the more complex FB-DIMMs found on the 5000 platform or the more scalable but perhaps still immature DDR3 alternative. The one notable change with memory is the improvement from the 3010's top bus speed of 667 MHz to the 3210's new 800 MHz. The architecture is still dual-channel, with a maximum of 8GB system memory using modules with up to 1 gigabit chips. (This is a small but important point as it rules out the use of some single-sided DIMM options.) Also note that the 3210-V offers a single x8 PCI Express connection while the 3210-P variant delivers either two x8s or a single x16 for more flexible add-in graphics processing. There is no integrated graphics support in any modern Intel workstation/server northbridge.

PLACED IN A PEDESTAL. Intel's SC5295-E is a tool-less, rugged chassis ideal for 1P servers that can quickly adapt with rails from pedestal orientation into a 5U rackmount deployment.

A single PCI-X segment is available in the 3210 platform, but the bigger spec in today's market is the six PCI Express lanes available off the ICH9R southbridge, one of which is dedicated to Intel's Gigabit Ethernet controller. This leaves another dedicated x1 connection, then a choice of either a single x4 connection or four more x1s. The ICH9R also provides six 3 Gb/s SATA ports (complete with RAID 0, 1, 5, and 10 support), 12 USB 2.0 ports, and PCI connectivity.

At present, there's not much to add about Garlow WS, the workstation platform based on the X38 chipset. Most likely, it will mimic the 975X workstation products from earlier this year in that there will be a select few motherboard SKUs much like the standard workstation models, only with more tweaking options for those who want to max out the speed of their systems. You could easily imagine this appealing to folks doing video rendering or complex simulations, for example, where every minute of processing time can mean significant dollars. Look for the X38 to pop up in places like Intel's X38ML (Melstone) motherboard, sporting a narrow format PCB suited to either blade deployment or a two-PCs-in-one-chassis design using Intel's 1U SR1520ML chassis.

In general, the Xeon 3000 family is suited to applications in which the most sensible solution revolves around assigning one or two multi-threaded tasks to a specific, cost-effective server box. Virtualization may have a minor role here, but it's better suited to the 2P and 4P platforms detailed below. Additionally, the 1P design makes perfect sense in situations where businesses can't afford for an application–even one with relatively low compute demands–to go down. You don't need maximum CPU cores here; you need affordable system redundancy. A 1P server fits the bill perfectly.

The 2P Sweet Spot: Xeon 5000/5100 Chipsets
We've dedicated a fair amount of ink to the Xeon 5000 (Bensley) platform in Tech Insight–and for good reason. Bensley is what put Intel back in the volume server limelight two years ago, and it remains the anchor of Intel's server line today, largely on the basis of its blend of performance, power efficiency, and affordability. In the channel, these attributes have translated into revenue opportunities that simply didn't exist before, both for basic components as well as the upsells that accompany them.

The last of Bensley's NetBurst-based Xeon 5000 parts–the 5050, 5060, and 5080–are being end-of-lifed as we write this. In their place, we have the dual-core 5100 and quad-core 5300 CPUs backed by the 5000X, 5000V, and 5000P chipsets. These three products are all very similar. The 5000X is the only member of the trio also noted as a workstation chipset. The 5000V supports a maximum of 32GB of FB-DIMM memory while the others can handle 64GB. Also, the 5000V omits one or two advanced memory functions, making it more suited to the "value server" category.

By and large, we can refer to these three 5000 models collectively as the 5000 chipset, and you can count on it staying current at least through the middle of next year. However, there's one major caveat to note: Late in Q3, the Bensley platform underwent a quiet refresh, and the only way you can tell if your motherboard uses a "pre" or "post" refresh version is to see if there is an "R" at the end of the motherboard's model name. If there is, then the board will support the dual-core 5200 and quad-core 5400 Xeon processors arriving this quarter. If not, then only the 5100 and 5300 series will work.

NICE PACKAGE. This image dates back to the Woodcrest (Xeon 5100) launch, but Intel's socket design has proven solid enough to survive a doubling of cores and two more Xeon generations.

Intel launched the 5100 chipset during the fourth quarter, coinciding with the launch of 45nm processors. One difference between the 5000 and 5100 chipsets is that the newer core logic is meant to be paired with the common ICH9R rather than the more specialized 631xESB/632xESB found with the current 5000. The differences between these two chips are minor. The ESB parts have integrated dual Gigabit Ethernet MACs, but the ICH9R offers integrated SATA RAID. The more important difference between the chipsets is the 5100's reversion from FB-DIMM memory to registered DIMM (RDIMM) DDR2. FB-DIMMs use considerably more power than their DDR2 counterparts, mostly because of their large buffer chips. Adopting DDR2 memory will allow Intel to demonstrate best-in-class performance per watt from the wall, while still delivering the more performance-oriented alternative with FB-DIMM.

Most recently, Intel launched its 5400 chipset for workstations. This marks a needed update to the 5000X chipset and the demands of today's quad-core CPU options. The 5400 raises the front side bus ceiling to 1600 GHz and supports both the 5200 and 5400 Xeon families. Also note that the maximum allowed FB-DIMM memory jumps to 128GB (up from 64GB) and now supports 800 MHz speeds rather than last generation's 667 MHz.

New Mainstream CPUs
We've already let the cat out of the bag on what you might expect from the latest Xeon 5000 series CPUs, but let's see what you can expect over the next couple of months. Throughout the third quarter, the 5100 series, topping out with the dual-core, 3.0 GHz, 80W model 5160, remained a hot ticket. Now, in Q4, look for the 5100 line to be gradually supplanted by the dual-core 5200 series, which uses its new 45nm fab process to create some impressive results.

Definitely pay attention to the main specs for each of these chips as you start making recommendations because they may not always be

BEHIND THE BOX. The Xeon 5300 and 5400 CPU lines are both quad-core, but the 5400's move to 45nm enables a faster FSB. Alternatively, the 5300 remains a value play at slower clock speeds.

what you expect. For example, all 5200 processors have 6MB of shared L2 cache (up from 4MB in the 5100 series–gotta love those smaller transistors) and a common architecture, but similarities end there. The E5205 starting point uses a 1066 MHz FSB, no doubt a key factor in why the chip draws only 65W while the other E52xx models draw 80W. The E5260 will hit 3.33 GHz on a 1333 MHz bus, and the E5272 will hold down 3.40 GHz on a 1666 MHz bus. Look at these just in terms of frequency and you'll see one part on the low end and two parts at the top–with nothing in the middle. Why? Probably because Intel presently has a large competitive advantage in manufacturing quad-core processors, and it plays to Intel's best interests to help nudge the market in that direction. So when dollars count, a small handful of dual-core parts may still be the best fit for a bid, but look to Intel to make quad-core options increasingly attractive.

Speaking of bids, there is one more 5200 part to note: the L5250. We know that many resellers were able to win business over the last two years with the Xeon 5148, Intel's 40W low-voltage (LV) part in the original Bensley lineup. Buyers would get most of the performance of a standard voltage part, and while the chip might cost a few dollars more, the month-to-month energy savings more than compensated. The L5250 continues that line of thought, now at 3.16 GHz on a 1333 MHz bus but still at only 40W. For any customer concerned about power bills or just eco-friendliness, the L5250 is a killer story.

Moving into the quad-core 5400 group, there are more SKUs but fewer surprises. Most of the family is based on a 1333 MHz FSB, and all have 12MB of L2 cache. The E54xx line starts with the E5405, a 2.00 GHz part that is the only member of the 5400 series to lack demand-based switching (DBS), Intel's processor feature that reduces a CPU's voltage and clock speed until an application requires more power. DBS is also missing from the L5250. Given that the L5250 is always running in a relatively low-power state (40W), this omission may make sense; it makes less sense in the 80W E5405. So be aware that the 80W E5405 may chew through considerably more power than the 80W E5410 (2.33 GHz), and the extra few dollars for the slightly faster part may yield more benefits than meet the eye.

WORK THAT 2P. Intel's S5400SF motherboard, sporting the new 5400 chipset, emphasizes memory scalability and high bandwidth. Look for this SKU within Intel's SR1560SF server systems.

The E5400 series ratchets steadily upward in frequency through the E5450 (80W, 3.00 GHz). Then we hit the X5460, a more performance-oriented SKU that keeps the 1333 MHz FSB but lands at 3.16 GHz only after taking a power jump up to 120W. Is it worth a 50% gain in the TDP spec in order to eke out a few more megahertz? Maybe on some time-critical applications, especially if Intel turns out to enable significant overclocking on this part when combined with the X38 chipset in workstation settings. For mainstream servers, though, we're guessing that the X5460 won't enjoy runaway volume sales. The E5462 (2.80 GHz, 80W, 1600 MHz FSB) and E5472 (3.00 GHz, 80W, 1600 MHz FSB) are too similar in performance, offer higher bus overhead, and don't incur the power penalty.

One last note on the 5400 line: With the L5250, the value proposition is clear. You get all of the performance at half of the wattage. Similarly, with the two LV SKUs in the 5400 line, the L5410 (2.33 GHz) and L5430 (2.66 GHz), the TDP takes a plunge down to 50W from the 80W of the standard "E" models. For the ultimate in performance-per-watt, these two SKUs are ideal.

MP: Core's Coda
Without question, the song of whitebox servers is more or less a two-verse strain: 1P and 2P. The MP segment (generally 4P up to 32P) is rarely served by channel system builders. Most of that business is handled by the tier-one OEMs as well as more specialized names such as Sun, NEC, and Fujitsu Siemens. In the whitebox world, MP systems account for well under 10% of the total server market. This explains some if not most of why Intel has given its Xeon 7000 family the least and last attention from its entire line of server offerings. The 7000 group is simply the lowest volume. That doesn't make it a bad product or one you shouldn't be ready to sell if and when the chance arises. After all, MP machines are good money makers if you can land them. But with finite resources and lots of battles to fight, the MP segment was last on Intel's priority list.

KEEPING COOL. The Advanced power efficiency of Intel's Core microarchitecture now applies in the MP space, enabling low-profile and even passive heatsinks in quieter, more cost-effective servers.

In late summer, it was time at last to retire the last NetBurst-based processors, the 90nm Xeon 70xx and 65nm Xeon 71xx series, and issue in Caneland, the platform name for Intel's dual-core Xeon 7200 and quad-core Xeon 7300 processors as well as the corresponding 7300 chipset.

With some server platforms, the difference between chipsets largely boils down to the number of processors they'll recognize. But Intel didn't let all of that wait time for Caneland boil down to counting cores. The 7300 chipset can handle up to four Xeon 7200/7300 processors on one motherboard. The FSB gets bumped up to 1066 MHz from the prior generation's 800 MHz. This may not sound like much compared to the 1333 MHz and 1600 MHz buses we saw earlier, but realize that Intel has revamped the dual independent bus structure. With today's desktop systems, there is one front-side bus pipeline between the CPU and the northbridge. With the 5000 series Xeon platform, there are two such pipelines, hence the name dual independent bus. With the 7300 chipset, Intel has increased this capability up to four pipelines, one for each processor, thus doubling the potential memory bandwidth. Intel now calls these Dedicated High Speed Interconnents. However, there will be a significant architecture shift with next year's Nehalem design.

With traditional Intel designs, managing multiple processors meant that Processor 1, for instance, would have to send out "snoop" queries through the northbridge and off to Processors 2 through 4. The repetitive latencies incurred by this snooping could slow overall system performance. So the 7300 chipset introduces a 64MB "snoop filter" designed expressly to optimize and accelerate these queries by caching a directory of current data and eliminating unnecessary queries.

A revamp of Intel's I/OAT also appears in the 7300, delivering performance improvements when used in conjunction with Intel's own NICs. The new I/OAT integrates low-latency interrupts, stateless interrupts, and MSI-X. This last item leverages multiple vectors, or queues, to minimize bottlenecks through improved load balancing across multiple cores. (The prior-gen MSI technology could only handle one processor.) When vectors are assigned to specific processors, only those cores receive interrupt requests, thereby reducing the traffic load on the other cores. The "next-generation improvements" in the 7300's I/OAT also include Direct Cache Access (DCA), which accelerates inbound I/O by sending it directly into a processor's cache.

Like the Bensley platform, the 7300 chipset adopts FB-DIMM memory at speeds up to 667 MHz, but this higher-end implementation supports up to 256GB per motherboard. Naturally, the chipset supports ECC for bit-level code errors, but Intel also adds features such as memory sparing, which can predict a failing memory module and copy out its data to other modules before failure causes data loss. This protection is further reinforced by memory mirroring, which is pretty much a RAID 1 for system memory that entails copying data to two locations so that uptime is preserved in case one fails.

Obviously, the more processors and memory you have in a box, the more suited it will be to virtualization. This is why Intel integrates a new feature called VT FlexMigration in the 7300. FlexMigration is a bed of hardware-based features designed to assist in live virtual machine copying and installation without having to shut down and reboot. Think of it as "hot transferring," just one more way to provide uninterrupted runtime for critical MP systems.

The 7200/7300 Family
By now, you know the pattern. The idea of a dual-core MP system is a bit odd in an age when quad-core is becoming the norm, so don't be surprised to find there are only two 7200 series SKUs. The E7210 is an 80W, 2.40 GHz part with 8MB of L3 cache. The E7220 sports the same specs, only at 2.93 GHz. Note that the 8MB of cache in these two models sits in an L3 repository, not L2 as in the other Core microarchitecture models. Last year's NetBurst-based Xeon 7100 series featured dual-core models with 4MB, 8MB, or 16MB of L3 cache. With the 7200 series settling on this middle amount, Intel would seem again to be making the 7200s neither too attractive to steal share from the quad-core units nor too lame to seem like an unworthy product.

While it is a minority player in the new MP lineup, keep in mind that the 7200 series has several advantages over the prior generation. The Core microarchitecture alone accounts for a considerable performance improvement. The 7100 series may have shown higher frequencies, but the added processing efficiencies of Core over NetBurst more than compensate. Additionally, the 7200s deliver this higher performance at nearly half the power: 80W versus 150W (models 7130N and higher). Not least of all, note that all present models in the 7200/7300 groups use a 1066 MHz front-side bus as opposed to the 7100 family's 667 or 800 MHz FSB. The 7200 models also offer a lower price per core than the 7300s. For example, as of October 2007, the 2.40 GHz E7210 shows up for $856 on Intel's price sheet. The 2.40 GHz E7340 doubles the core count but more than doubles the price at $1,980. Of course, by the time you make it to MP server price ranges, a few hundred bucks more for chips that enable higher compute density is a drop in the bucket. On the other hand, a researcher wanting a modestly priced compute "cluster" could take a 4P board using Xeon 7200 processors and have what not so long ago would have been called an 8-node cluster sitting under his desk.

BACK TO THE FUTURE. While the single-core Xeon 7140M may have boasted a 3.4 GHz clock, the quad-core, Core-based X7350 (2.93 GHz) blows away its predecessor across many performance tests.

The quad-core 7300s start with the E7310, a 1.60W part with 4MB of L2 cache per core. And here's the interesting choice: Intel's October pricing for the E7310 is $856, the exact same price for the lower-end 7200 model, the E7210. So which do you pick? The dual-core with the higher clock speed or the quad-core with the low clock speed? Good question. We suspect the quad will be the most frequent choice because it offers way more thread scalability at the same price point. If clock speed were the higher priority, the quad-core, 2.66 GHz Xeon X5355 delivers twice the cache and chimes in for less green at $744. Naturally, these prices will all change in the coming months, but you get a sense of Intel's server positioning and how various server chip features will influence the price stack. With the 2.40 GHz E7330, Intel bumps up the processor's total L2 cache to 6MB (2x3MB) at a price point fairly near the 2.13 GHz, 4MB L2 Xeon E7320 right below it. The next step up is the E7340, exactly like the E7330 save for 8MB (2x4MB) of L2 and a nearly $600 price delta. So unless your customer's application(s) are very sensitive to memory caching––and many MP-optimized apps are–the E7330 may be the bargain of the bunch.

The two remaining SKUs in the 7300 line are the L7345 and the X7350. The oddly numbered L7345 is the family's stand-alone low voltage part, hitting home with a surprisingly cool 50W TDP. There's some frequency trade-off involved, of course. The LV quad-core chip only runs at 1.86 GHz, so again, the emphasis here would be on power savings and maximum thread counts rather than high frequency rate.

Going the other way, the X7350 zooms up to 2.93 GHz, still keeping the 8MB of L2 cache and 1066 MHz front side bus. With the blistering speed firing across four cores comes a bump in TDP up to 130W. Obviously, this isn't your first choice for an eco-sensitive bid, but it is your best option when application processing speed is paramount. Also note that the X7350 and L7345 carry the same sticker price at the top of the Xeon cost ladder. The message is clear: In the MP world, there are two ways to get the top chip on the market. You can pick the lowest power envelope or the highest performance envelope. Both attributes carry the same value depending on the customer's needs.

MEET THE S7000FC4UR. This 4U, 4P Intel server platform is based on the 7300 chipset and features data center-friendly hot-swapping, redundancy, and manageability throughout the system.

That's a lot of speeds and feeds to digest, but if you can handle just a few more numbers, here's the real take-away: According to Intel, the architectural advances in the 7300 series deliver up to a 1.86X performance improvement over the quad-core Xeon 5300 line, only some of which results from a quadrupling of platform memory capacity. For customers who might still be looking to save a few bucks by going with the older, NetBurst-based 7100 line, remind them that the 7300 family offers up to a 3X performance-per-watt advantage. All told, the 7300 line generally delivers a 1.5X to 2X improvement in applications over counterpart members of the 7100 series.

Naturally, the CPUs may garner all the glory from these performance gains, but behind the scenes the new 7300 chipset is pulling its own weight and making the 7300 processors' new paper specs into real world gains. This chipset is what enables that FSB hop up to 1066 MHz and brings the Xeon platform current into Core microarchitecture compatibility. This is where Intel implemented its new Dedicated High-Speed Interconnects (DHSI) architecture, now providing a direct point-to-point link between each processor core and the northbridge. This step alone doubles the platform's memory bandwidth. The northbridge provides both PCI Express support as well as optional platform inclusion of Intel's 6700PXH 64-bit PCI hub, which offers twin 133 MHz PCI-X segments, each with its own hot-plug controller.


The Need for Xeon
We should point out that even in the dark days of 2003 and 2004, there was a reason Xeon processors hung onto their market dominance and maintained millions of loyal customers. For most server buyers, everything we've discussed above–frequencies, TDPs, core counts, and all the rest–is only one consideration when settling on a server platform.

PILE ON THE OPPURTUNITY. Whether your interest in Xeon applies to the chips, the platform, or even stacked rackmount arrays, today's SKU diversity delivers unprecedented power and value.

Another primary concern is stability. A big element there is being able to have the processor, chipset, motherboard, and even chassis come designed, manufactured, validated, integrated, shipped, and supported from one common source. Then as now, Intel is the only source in the industry able to claim such a unified production chain, and no one comes close to matching Intel's renowned channel support on server products.

"With over 2,000 integrators and builders building quad core Intel Xeon processors already in 2007," says Intel's North American channel products marketing manager, Todd Garrigues, "we're very excited about these latest releases and the extraordinary value that you can provide to previous customers. This gives businesses good reasons to upgrade their infrastructure. Our goal is to have 80% of our customers selling quad-core server processors in the next six months, but that's only one part of the whole Intel server picture. With our servers, a customer can get boards, processors, cases, and other peripherals from one company with one support phone number and one set of benefits and resources. One of our biggest channel goals for the coming year is to be easier for our customers to do business with."

There are server applications from humble mail machines to critical Fortune 500 database systems, and now Intel has a full range of Xeon platforms, all based on the industry-leading Core microarchitecture, ready to tackle each of them in turn. You should now have a solid sense of the Xeon platform lineup as we head into 2008. The next step is to set these processor and chipset foundations before your clients and determine how they can best make your customers' daily operations more effective and profitable.