RAM | Reseller Advocate Magazine: Issue 58

By Chris Angelini

The last time we talked workstation hardware, Intel had just released Bensley, ATI was launching its first Shader Model 3.0 professional card, and NVIDIA had sent us a sample of its ultra high-end Quadro FX 5500 monster—and this was only a few months ago. Nevertheless, quite a bit has happened in the professional space since then. AMD recently revamped its Opteron processor, changing the chip's socket interface for the sake of extra performance. NVIDIA invested further into its server and workstation lineups too, augmenting the capabilities of the Quadro and nForce Professional families. With so many changes in such a brief span of time, savvy resellers should have little trouble uncovering opportunities in the workstation market.

Opteron: Revised and Revisited

AMD's new Opteron represents one of the most significant landscape changes, breaking pin compatibility with the Socket 940 interface. Resellers have relied on that socket interface since Opteron first emerged in 2003, persisting even through the move from single- to dual-core processing. To adopt something new, AMD must have really made some noteworthy changes to its core.

Indeed, the reworked professional chip bears a handful of improvements that really push toward more modern standards and features. Foremost on the list is an advanced memory controller equipped to accommodate DDR2 modules. When AMD first launched the Opteron, its inclusion of onboard memory control logic gave it a massive performance boost versus competing processor designs reliant on a northbridge for similar requests. The clear caveat was that any progression in memory technology would take longer to integrate than it would in a northbridge-based solution. Nevertheless, super-aggressive memory settings and admirable scaling validated AMD's vision; Opteron dominated the benchmark charts for three years as a result.

Throughout the processor's first few years, AMD maintained its design would work best mated to DDR; the DDR2 variants available sacrificed too much timing in the name of clock speed. Now, however, DDR2 is selling at speeds in excess of 1 GHz, delivering the frequency needed to overcome timing deficiencies inherent to the technology. When you consider that the roadmap for traditional DDR is starting to wane, it makes sense for the company to shift focus.

Opteron processors with DDR2 support will be available in three flavors: the 1000-series, 2000-series, and 8000-series. Like their predecessors, the new chips use model numbers to convey performance, though they've moved from three digits to four. The dual-core 1000-series chips actually plug into Socket AM2 motherboards and work with DDR2-800 memory, just like the latest desktop Athlon 64s. They only operate in single-processor configurations, though. Stepping up to a 2000- (2-way) or 8000-series (up to 8-way) Opteron necessitates a motherboard with Socket F (1207-pin) compatibility. Those models need registered memory modules, too, which means adopting slightly slower DDR2-667. As a side-effect of switching to DDR2, the latest Opteron platforms consume less power than their predecessors, giving your customers yet another reason to upgrade.

But the new socket interface isn't just about saving power or delivering marginal performance increases. Rather, AMD claims that it's laying a foundation for quad-core processors due out in 2007, just as Socket 940 facilitated a seamless shift from single- to dual-core over the past year. Despite the planned increase in computing complexity, AMD is confident that its quad-core solutions will fit within the same power envelope as its existing dual-core design manufactured at 90nm.

The final feature added to AMD's revised Opteron is hardware-based virtualization, dubbed AMD-V, matching a similar capability Intel built into Xeon. More than anything, integrating virtualization is a matter of good timing for AMD, since Intel is starting to really evangelize the feature.

Everything else about the Opteron is, for the most part, as it was. From the HyperTransport interconnects to the 64-bit extensions and PowerNow! technology, all of the benefits that made Opteron so popular in the first place persist with the latest core revision.

nForce Professional:
Workstation Logic

Of course, new processor interfaces are generally accompanied by cutting-edge platforms. And since AMD is emphatic that it isn't in the chipset business, it relies on chipset partners—ATI, Broadcom, and NVIDIA—to deliver the goods. Despite AMD's recent announcement of plans to acquire ATI, the Canadian firm has yet to dabble in professional chipsets. NVIDIA, on the other hand, is a seasoned veteran. Its nForce Professional solutions have long driven single-processor workstations all the way up to 8-way rackmount machines. A fresh trio of nForce configurations now fill out the company's lineup, timed almost perfectly to complement AMD's Socket F unveiling.

The most impressive platform consists of two chips, the nForce Professional 3600 and 3050. Combined, the pair proffers an astounding 56 lanes of PCI Express connectivity across 12 possible links. Fully exploited, the chipset also enables 12 SATA drives, four PATA drives, and onboard RAID functionality. The mind-blowing stats don't stop there—Gigabit Ethernet is natively supported over no less than four onboard MACs. The quartet is load balanced in hardware and protected through a fail-over mechanism. Should one MAC go down, the others provide redundancy. In a bid to reduce CPU utilization while processing network traffic at such lofty speeds, NVIDIA also accelerates TCP/IP in hardware.

If you think such a setup sounds like overkill, take a look at Tyan's Thunder n6650W motherboard, which employs the 3600/3050 combination. The eATX board comes ready for a pair of 2000-series Opterons. Eight memory slots handle up to 32GB of DDR2-667 expertly. And four PCI Express x16 slots (two wired to run at x16 and the others at x8) open the door to SLI rendering, high-end SAS, 10 Gb Ethernet, or any other combination of powerful peripheral cards. High-definition audio, ten USB 2.0 ports, six SATA 3 Gbps connectors with onboard RAID 5, and dual Gigabit Ethernet controllers round out Tyan's feature-fest. Although the Thunder n6650W doesn't dish out its chipset's peak functionality, the board is remarkably robust and priced well under $500.

The nForce Professional 3600 also works by itself on motherboards with a bit less connectivity. Shedding NVIDIA's 3050 companion chip leaves 28 lanes of PCI Express, six SATA ports, and two Gigabit MACs. TCP/IP acceleration, load balancing, enclosure management, and SLI are all still included. ASUS' KFN5-D solidly illustrates the platform's potential, exposing two PCI Express x16 slots (one of which runs at x8), a total of eight SATA 3 Gbps ports, high-def audio, and dual Gigabit controllers. Priced just above $300, the KFN5-D makes a great workstation board, even if its PCI Express subsystem isn't as well-endowed.

NVIDIA's nForce Professional 3400 has all of the same specs as the nForce Professional 3400, only its PCI Express lanes are not configurable. While acceptable in lower-end graphics workstations, most resellers will want to stick with the more flexible 3600 as a means to provide flexibility.

Quadro Plex 1000:
Shaking Up Graphics

There's a clear pattern in NVIDIA's most recent chipset releases: I/O above all else. Where 40 lanes of PCI Express were once standard, 56 lanes now rule. Why offer eight SATA ports when you can expand that to 12? If your customers can find a way to use two Gigabit controllers, give them four and see what happens.

SLI multi-rendering technology is supported across the company's entire line of professional chipsets, in many cases through dual x16 links. That combination opens the door to some interesting capabilities unavailable from any other vendor. Recognizing opportunity, NVIDIA came up with the Quadro Plex 1000—a visual computing system akin to what your customers might have once seen from Silicon Graphics or Evans and Sutherland, according to Jeff Brown, general manager of NVIDIA's professional solutions group. Housed externally, the Quadro Plex either stands upright or mounts in a standard 19" rack. It's completely self-contained, sporting an independent 480W power supply, cooling subsystem, and proprietary connector used to interface with a host workstation.

"There are three different versions of the Quadro Plex 1000, each targeted at different vertical markets," says NVIDIA's Brown.

Model I features two Quadro FX 5500 cards in SLI, each with 1GB of memory—ideal for applications heavily dependant on 3D performance. Oil and gas software, for example, processes large quantities of volumetric data and are able to benefit from the massive capabilities of two ultra high-end cards working cooperatively. Model II includes two Quadro FX 4500 X2 boards—two cards, each with two GPUS. A total of eight dual-link DVI outputs make that version perfect for digital signage, for instance. Moreover, support for up to 64x SLI anti-aliasing endears the setup to engineering apps, which rely on straight lines and properly represented pixels. Model III also sports two Quadro FX 5500 boards but with an SDI card included to output 8-, 10-, or 12-bit uncompressed video. Without question, that configuration is intended for use in broadcast video applications handling high definition content. "We're able to leverage our programmable GPU architecture to manipulate pixels in every way imaginable, including the pixels you see in video. Quadro Plex, complemented by the SDI card, enables fully configurable video channels, color space conversion, and gamma correction. When you watch a football game and see the computer-generated first-down line across the field or on-screen diagrams, that's all done with similar broadcast graphics systems."

NVIDIA's Jeff Brown adds that selling such specialized products demands the expertise of an educated cadre of channel partners, since the least expensive Quadro Plex 1000 starts at just over $17,000. Pricey though they may be, though, each model is able to interface with any system sporting a PCI Express x16 slot (NVIDIA plans to certify platforms for Quadro Plex support). SLI motherboards may even accommodate two enclosures. Windows XP and Linux are both supported in 32- and 64-bit trims. Finally, NVIDIA supports the trio through its UDA (Unified Driver Architecture), so there's no need to worry about special software or lacking technical help for resellers.

A Landscape of Choice

Clearly, it's a good time to be involved with workstation hardware, regardless of whether you ally with Intel, AMD, NVIDIA, or ATI. The Bensley platform launch propelled Intel out of its defensive position, and now AMD's DDR2-equipped Opteron is evening out the playing field once again. On the graphics side, NVIDIA is using SLI to extend its advantage in the professional space since ATI's CrossFire is still exclusive to enthusiast desktop graphics. As ATI focuses its attention on claiming single-card market share from the mainstream to ultra high-end, NVIDIA's nForce Professional and Quadro Plex products demonstrate the flexibility of SLI—the only multi-card workstation solution—beyond gaming.