Take chipsets as an example. At one time, AMD swore off the core logic business, preferring to focus on its processors and let third parties control the silicon linking hardware such as graphics cards and storage to the CPU. With the acquisition of ATI, however, AMD tapped into that company’s growing chipset expertise and turned the channel onto a message Intel popularized a long time ago: Platforms are more valuable than individual components.

AMD’s work with platforms has made it much more attractive in the channel. The GAME! initiative, to begin with, is an interesting play on simplifying entertainment PC purchases to the point of mimicking the way gamers buy consoles. LIVE! sets the stage for a richer home theater environment, bolstered by a reseller-ready software foundation developed by AMD’s own team. And then there’s Business Class, a nod in the direction of VARs looking to add value by offering their SMB customers persistent hardware configurations. Not only are the Business Class platforms designed for a long useful life, but they’re also easily serviceable, efficient, and more secure.

A NATURAL FIT

For all of AMD’s efforts in getting its mobile and desktop processors into well thought-out platforms complemented by attractive home-grown chipsets, there remains one market where the company’s processors speak for themselves, take the help of third-party core logic, yet remain viciously competitive against Intel’s best efforts: the multi-processor space

We’re not talking about multi-core here. The best Opterons already boast four cores apiece, and those are the ones resellers invariably want to be selling—and for good reason. It’s much less costly to build a 1P box with one quad-core chip than a 4P system utilizing four single-core processors. When that’s the sort of processing horsepower your customer needs, taking the path of least complexity saves precious IT dollars.

Value in the 8-SerieS: If your customer needs a box with more than four processors, the 8300-series Opterons have what it takes to enable four- and eight-way systems.

But what about the businesses that can’t get enough muscle; the companies that were buying four- and eight-way machines back when processors only offered one core each? Today, those same organizations still have access to 4P and 8P server hardware. Only now, those same systems wield between 16 and 32 cores, along with many, many gigabytes of system memory. This is where AMD’s Opteron really shines.

Perhaps you’re wondering why AMD’s fastest Phenoms stake out value positions against Intel’s mid-range quad-core contenders while Opteron blows to the front of the line in higher-end applications, right up to the high-performance computing (HPC) market. The answer has a lot to do with architectural decisions made by the minds behind AMD’s newest K10 design.

If you look at many benchmarks, you’ll notice that games have a tendency to favor processors with plenty of fast cache memory. For instance, CPUs based on Intel’s 45nm Yorkfield include a total of 12MB of L2 cache, divided into two 6MB shared repositories. In contrast, the K10 microarchitecture features half a megabyte of L2 per core, along with a shared 2MB L3 cache. Given the high latency of K10’s L3, it’s really no wonder Intel consistently does so well in most desktop applications.

Shift your focus to the server and HPC world, where large, heavily threaded workloads put a real squeeze on memory performance and dedicated L2 caches enjoy more bandwidth than the shared storage space implemented by Intel. Suddenly, K10 makes a lot more sense. And don’t think the processor’s benefits are limited to its execution cores or cache implementation. Opteron’s integrated memory controller is also a tremendous boon, as is its HyperTransport interconnect.

An Architecture for Multi-Processing

AMD still sells its dual-core Opterons, in addition to the newer quad-core versions centering on the Barcelona and Budapest cores. But when it comes to value, there’s no question that quads are the way to go.

Opterons destined for 1P servers fall under the 1300-series model designation. The chips are manufactured on a 65nm node and drop into a Socket AM2+ interface. They’re the first Opteron processors to incorporate HyperTransport 3.0 support and a DDR2 memory controller able to drive 800 MHz modules. Although the 1300s are officially based on the Budapest design, you’ll notice a lot of similarities between the 1P Opterons and AMD’s Phenom X4. These chips are AMD’s equivalent of Intel’s Xeon 3200 family—desktop in nature, but optimized to run on business motherboards..

The Opteron 2300-series changes that infrastructure dramatically. Socket F replaces the desktop AM2+ interface, adding support for a second physical processor and yielding up to eight cores in a box. No longer derived from Budapest, the 2300’s Barcelona core is a bit older. It’s limited to 1 GHz HyperTransport speeds and employs registered DDR2 memory rather than unbuffered modules. Of course, the more enterprise-class memory technology is a little less aggressive, topping out at 667 MHz (533 MHz if you want to use all eight memory DIMMs supported per CPU).

Native Excels in the EnterprisE: AMD’s K10 microarchitecture might be outgunned by Intel’s Core 2 on the desktop, but its native quad-core design really takes off when you shift to servers.

AMD’s most potent quad-core Opteron processors belong to the 8300-series, which enables 4P and 8P configurations. This is where the rubber really meets the road—a market where AMD has chewed away at Intel’s share using a more elegant implementation. The 8300s are also Barcelona-based, meaning they utilize the same Socket F interface, 1 GHz HyperTransport frequency, and 667 MHz DDR2 memory controller as the 2300s. However, those architectural features serve the 8300s even more dutifully given the complexity of a multi-socket platform and potential for bottlenecks.

Although Intel’s competing Xeon MP (employing the Tigerton core) is now backed by dedicated high-speed interconnects to each processor, the front side bus topology is still at a disadvantage to AMD’s HyperTransport technology. The 7300 chipset’s use of FB-DIMM memory does allow it access to tons of memory, but with a caveat. FB-DIMMs are already viewed critically for their elevated power consumption. Dropping 16 or 24 modules into a 4P server is really going to impact that machine’s energy efficiency story.

Chasing the Green Angle

In comparison, the 8300-series Opterons include an integrated DDR2 memory controller. Not only do DDR2 modules use a lot less power, they’re also significantly more affordable than FB-DIMM memory. You can expect to pay less than $150 for 4GB of registered DDR2-667. An FB-DIMM kit in the same capacity, running at the same speed, costs $100 more.

Plan a Scalable Platform: Build an SMB’s next server using the Tyan Thunder n4250QE motherboard and a 5U chassis. That way, if the customer ever needs to upgrade, you can add up to four more quad-core CPUs.

According to Burke Banga, business development manager for the Opteron processor in North America, there’s also a compelling story to tell customers when it comes to CPU-specific power optimizations. Built-in Enhanced PowerNow! Technology facilitates dynamic voltage and frequency adjustments based on processor loading, minimizing consumption during idle periods. Especially as you move from 1P and 2P configurations and into the 4P and 8P systems, the energy savings can get quite granular. The Dual Dynamic Power Management feature allows the CPU and memory control to be tweaked independently. From there, individual core utilization is measured and, if one or more cores aren’t being taxed, they can be spun down to zero in on power savings and reduce heat output. Finally, AMD employs clock gating to turn off portions of each quad-core chip not in use.
.
Between AMD’s energy optimizations and the benefits inherent to standard DDR2 memory, we’d argue that there’s an exciting play on maximizing the performance your customers are able to extract per watt of power consumed. This is a particularly important point for SMBs setting up a handful of servers without having to pay extra for a reinforced power infrastructure.


Picking Powerful Platforms

Right now, AMD’s platform demands are being met by NVIDIA and Broadcom. NVIDIA’s offering includes the nForce Professional family, while Broadcom sells the HT-2000 and HT-2100 SystemI/O Hubs. Of course, everyone is waiting to see a chipset emerge from the union with ATI. And while AMD confirms that the core logic is in development (and hotly anticipated, given the strength of a complete server platform story), the company isn’t quite ready to talk details. We were able to dig up a roadmap intended for financial analysts covering AMD, though, projecting a mid-2009 release of the company’s first server chipsets. Fortunately, both of AMD’s existing partners offer robust chipsets available in some incredibly inventive platforms.

Piggybacking to 8P: Tyan’s M4985 daughter card expands the Thunder n4250QE’s processor handling from four processors to eight, doubling memory capacity in the process.

Let’s start with the juggernaut: Tyan’s Thunder n4250QE, a four-way Opteron board with an interesting upgrade story. The Thunder employs NVIDIA’s nForce Professional 2200 and 2050 components to enable four PCI Express x16 slots (two of which run at x4 signaling rates), eight SATA 3 Gb/s ports, three Gigabit Ethernet controllers, and USB 2.0 connectivity. Each of the four 1207-pin Socket F interfaces takes a dual- or quad-core Opteron 8000-series processor. And four memory slots attached to each socket take a total of 64GB of DDR2-667. Now, in the center of that foursome you’ll find a pair of slot connectors, both of which are HyperTransport interfaces. Tyan’s M4985 daughter board drops onto the Thunder n4250QE’s HyperTransport connectors, offering four more sockets and 16 additional DDR2 memory sockets in a form factor you can tuck into a 5U chassis. Add it all up and you’re looking at the potential for 32 processing cores and 128GB of memory. That’s staggering computing horsepower.

Need even more muscle? Supermicro’s SuperBlade infrastructure lets you build up density as your customer demands it. The SBE-710E enclosure fits as many as 10 hot-plug processor blades in a 7U box. But this isn’t just a play on saving space. By consolidating to a blade enclosure, you also contain the cabling requirements of 10 machines, exposing only a handful of necessary cables such as power and Ethernet connections.

Petite Processing HorsepoweR: Take advantage of 16 cores in a single blade module with Supermicro’s SBA-7141M-T, designed exclusively for high-performance computing.

Supermicro offers a couple of different Opteron-based options, including one four-way and one two-way blade. Of course, if you’re building the densest systems possible, 10 four-way blades are the way to go. The SBA-7141M-T is the ultimate tool for maximizing computing horsepower per 1U of rack space. Each diminutive blade boasts an NVIDIA nForce Professional chipset, four Socket F interfaces, and 16 240-pin DDR2 memory slots. A single SATA port and enough space for one 2.5” drive per blade is indication enough that this is no storage server. Rather, it’s all about AMD’s Opteron processors. And in the interest of moving data from node to node quickly, each blade also sports two InfiniBand slots. Supermicro’s InfiniBand switch module offers 10 4x DDR ports, each able to move 20 Gb/s.


Where Does the Reseller Fit In?
Knowing a bit more about AMD’s flagship Opteron 8300-series server processors and the platforms that drive them, you might be wondering if there’s really a channel opportunity here. After all, wouldn’t it make sense for a medium-sized business or enterprise to go with a tier-one for big iron purchases like eight-way servers?.

AMD’s Buke Banga says that’s actually not the case. “We like to talk about six different deals from the Barcelona launch. Of those, three are tier-one deals, and the other half are from channel partners.” Remember back to 2003, when AMD started shipping the very first Opterons; it had only managed to land IBM. So, the system builder adoption this time around really says a lot about the market’s acceptance of AMD’s server CPU.

More Bling, Less WorK: With an enthusiast-minded tower from Thermaltake, a lot of the extras already come included, saving everyone time and expense after the sale. Units like this Armor MX feature loads of expansion room

Banga says that VARs can do things with the four- and eight-way processor not possible through tier-ones. Customizing power delivery is one example. Or you can incorporate InfiniBand. Because AMD has won over so many channel partners, there are plenty of options available to the reseller looking to differentiate, and that simply wasn’t the case a few years back.


Moving on Up
Once upon a time, motherboard vendors hid the fact that they supported AMD’s Opteron CPUs for fear of reprisal from Intel. Today you’ll find more Opteron-based 1P, 2P, and 4P platforms than ever before. You’ll also find 8P platforms—something that can’t be said for systems centering Intel’s Xeon MP. A handful of chipsets help drive the dual- and quad-core AMD processors. And a contingent of channel partners has made the infrastructure available to VARs, from chasses to motherboards to networking to memory.


There’s even more good news for resellers exploring four- and eight-way Opterons right now. According to AMD’s Burke Banga, the next generation Opteron processors will leverage Socket F interfaces being used today. That takes us through 2009 with the Shanghai design. Shanghai is expected to be fairly similar to Barcelona. It’ll incorporate a larger 6MB L3 cache and core-level enhancements to improve the number of instructions per cycle processed. Those improvements are being made possible by a shift to 45nm manufacturing, at long last.

Further into 2009, AMD will shake things up once again by debuting Montreal with up to eight cores per socket, 1MB of L2 per core, as much as 12MB of L3, DDR3 memory support, and four HyperTransport 3.0 links for even more system bandwidth. Also etched on a 45nm process, that’ll likely be the first time we see AMD server chipsets. And of course, there’ll be a new socket interface to replace the venerable Socket F. Until then, resellers have a wide open opportunity to sell a processor architecture designed to excel in the enterprise space, backed by a stable socket interface, and developed into a rich infrastructure by channel partners eager to help VARs differentiate.