RAM | Reseller Advocate Magazine: Issue 64

By Chris Angelini

Over the years, I've done a lot of benchmark testing on overclocked components, and some people would say that makes me an overclocking expert. I know gamers who take a lot of pride in water cooling their systems and running system monitors tied to blinged-out, rheostat-studded bay devices. Friends of these gamers consider them overclocking experts. But the reality is that I actually know enough about overclocking to appreciate that I'm not an expert. The dozens of options you can tweak just on memory timings in enthusiast board BIOSes leave me baffled. I can fumble my way through, but I'm a generalist at best. Many resellers I've spoken with are in the same boat, which is unfortunate.

Microsoft's Windows Vista operating system is an expensive piece of software, whether you're buying the Ultimate version, one of the two Home packages, or the Business edition. And although many SMBs will want to wait six months or a year before they start experimenting with the much-delayed OS in mission-critical environments, the fact that many boxes are coming with Vista pre-installed or bundled with a coupon for a free upgrade means the software is on course to replace XP and is picking up momentum.

As you already know, a lot of Microsoft's attention in writing Vista was devoted to graphics. The result is a brand new user interface, which requires a certain degree of advanced 3D processing. Sure, you can still run Vista on an older box with an integrated graphics processor, but the experience will leave your customers wondering why they spent so much money on new software similar in appearance to Windows XP. Maximizing a Vista installation means springing for a premium GPU.

Of course, there's more to Vista graphics than its Aero glass interface. DirectX 10 is also a significant piece of the operating system that you won't get in XP. Based on an entirely new driver model, DirectX 10 introduces performance improvements to dynamic link libraries and nixes support for older versions, which played a role in slowing DirectX 9 down. One of the most notable DirectX 10 components is the Shader Model 4 specification, a high-level language that opens the door to some amazingly realistic effects. Naturally, if you want to see those effects, you'll need a DirectX 10 application and the hardware capable of driving it.

Hold Your Horses

According to former Dell CEO Kevin Rollins, the company sold tens of thousands of copies of Vista in the first weekend it was available. As promising as that might sound to resellers hoping to move lots of Vista-ready graphics cards, Danny Shapiro, senior marketing manager of workstation graphics at AMD, advises the channel to delineate between mainstream consumers and professional customers, exercising special caution with the latter group.

"The consumer—or gaming-oriented—segment is the one able to go for instant adoption. But when you have a new operating system, the customers using mission-critical business applications are not in a hurry to upgrade. Some of that is the logistics that have to be worked out across an organization. But an even bigger part is the differentiation that makes a workstation product a workstation product, from testing to certifying with Microsoft to working with the ISVs. Many of the third-party developers are busy working on their own software packages, and it'll probably be Q2 before they release versions that take advantage of Vista's new features. We're targeting the same time frame for certified drivers."

In other words, customers using professional graphics solutions might have more to lose from a quick Vista migration than they'd gain, prompting many to take a "wait and see" approach. Additionally, while talking to Shapiro, it also became clear that it's still a bit early to start upselling the DirectX 10 component of Vista—at least as the API pertains to professional graphics.

"In reality, most professional applications today are based on OpenGL. DirectX 10 really isn't a huge factor yet. Autodesk has a few applications that are based on DirectX, but even 3ds Max has the option to utilize either DirectX or OpenGL. With OpenGL your customer has the option to run on multiple platforms, and some ISVs want to maintain that support."

While it may be true that there isn't a huge contingent of professional DirectX-based graphics apps, Shapiro acknowledges that many are closely evaluating the API. And although Autodesk's popular 3ds Max does extend support for the DirectX and OpenGL APIs, company representatives are more than happy to talk about the upcoming 3ds Max 9 Extension: Productivity Booster package, which will give the software suite full Windows Vista and DirectX 10 compatibility. Available by the end of March and downloadable by any Autodesk Subscription customer, the update should help showcase some of what DirectX 10 can do.

We're Getting There

Even as your business customers sit on the sidelines, waiting to judge Vista on its stability, security, and software compatibility, both AMD and NVIDIA are furiously working on the hardware that'll drive the next generation of Vista workstations.

"After all," says AMD's Shapiro, "people will be moving to Vista. A lot of man-hours went into optimizing the operating system's performance and security. Once the professional graphics folks are ready to take the leap, we want to be ready with the cards they'll use."

In the meantime, competitor NVIDIA is already talking about the ultra-high-end cards ready for SMBs making the jump to Vista. Its Quadro FX 5600 is the new flagship pro board, ringing in with a $2,999 MSRP, and the Quadro FX 4600 falls in right below it for $1,000 less. Both offerings center on NVIDIA's G80GL graphics processor, closely related to the G80 that drives the GeForce 8800-series.

The G80GL represents NVIDIA's brand-new professional GPU architecture, and it's like nothing the company has ever offered before. Compared to AMD's current FireGL lineup, entirely comprised of DirectX 9 cards that'll be fully Vista-compliant given driver development time, the G80GL is a pure DirectX 10 processor with the right compatibilities to ensure an optimal experience in older applications as well. In order to meet Microsoft's DirectX 10 specification, which sheds the optional capabilities that convoluted DirectX 9, NVIDIA adopted a highly efficient unified shader architecture. Instead of implementing separate vertex and pixel shading engines, NVIDIA crafted an extremely parallelized core loaded with 128 stream processors organized into eight clusters. Understandably, the chip is incredibly complex, requiring 681 million transistors to build. Great performance is the payback for NVIDIA's four-year investment in designing the GPU. According to NVIDIA, math-heavy shaders run up to 4.4 times faster than on the preceding Quadro generation, high dynamic range lighting effects are 4.6 times faster, and scenes with heavy anisotropic filtering run up to 4.2 times quicker. When it comes to productivity-per-hour, that's serious improvement.

Both cards based on the G80GL are as sophisticated as the chip they host. The flagship Quadro FX 5600 is a full 12" long and two expansion slots thick. Because it's rated at up to 220W, the 5600 requires two auxiliary six-pin power connectors. Obviously, the board begs to be put into a roomy workstation with plenty of circulation and a hefty power supply. NVIDIA loads its Quadro FX 5600 with 1.5GB of memory, which feeds the graphics processor at 76.8 GBps. Two HDCP-protected dual-link DVI outputs grace the board's back bracket, driven by an all-new I/O companion chip that handles SLI processing and the display connectors. Given its price tag, the Quadro FX 5600 is considered an ultra-high-end card tailored specifically to customers manipulating large datasets. The usual suspects—oil, gas, and medical imaging—fall in at the top of NVIDIA's list of target markets.

The Quadro FX 4600 gravitates around the same GPU architecture, though the board is only nine inches long and, thanks to a 150W thermal rating, powered by a single auxiliary connector. Still, it's no slouch. The FX 4600, populating two expansion slots, wields 768MB of RAM pumping 67.2 GBps to the GPU. The same I/O chip drives a pair of dual-link DVI outputs. Your customer will pay less for a Quadro FX 4600 because it isn't quite as fast.

In all aspects other than rendering speed, the Quadro FX 5600 and 4600 cards are very much alike. They both support SLI multi-card rendering. The pair is also compatible with NVIDIA's existing Genlock and SDI add-ons for customers involved in broadcast media and high-definition content creation. Both new Quadros offer 10-bit color output, yielding the more accurate images valued by users in fields like medicine and design. The same folks will enjoy incredibly fast OpenGL performance, which goes hand-in-hand with big performance in DirectX and applications optimized for Shader Model 4.0.

A New Age of Multi-Processing

The G80GL graphics processor at the heart of NVIDIA's Quadro FX 5600 and Quadro FX 4600 professional cards is an extremely concurrent, data-centric computing engine. Similar to Intel's Xeon or AMD's Opteron in that it can handle multiple threads simultaneously, the chip's unified shader architecture introduces the concept of leveraging many (up to 128) floating-point processors able to operate on different types of data at the same time.

The idea of offloading certain types of computational tasks to idle graphics processors has been discussed by both ATI and NVIDIA. And although it doesn't pertain specifically to the relationship between professional graphics and Windows Vista, NVIDIA is introducing its CUDA initiative alongside the new Vista-ready Quadros, which may lead to some powerful performance improvements for customers who go ahead and adopt one of the pro cards in a Vista environment.

In essence, CUDA is a programming architecture that enables graphics processors to solve problems in business and technical applications. In the types of software most sensitive to multi-threading, such as physics, wave equations, matrix numerics, and finance apps, the throughput of NVIDIA's 128 stream processors opens the door to more speed than what a developer could get from any general purpose CPU. Previous attempts at GPU computing employed the OpenGL and DirectX APIs to limited success. CUDA employs a more suitable C-like programming interface that NVIDIA hopes will attract software developers. Customers eyeing CUDA should remember that only the latest generation of unified shader cards support the software interface.

Proceed with Caution

NVIDIA's first Windows Vista WHQL-certified driver emerged almost a month after Vista launched, and it isn't perfect yet. AMD is still shipping its RTM driver, which doesn't have the software certifications that make professional graphics cards more valuable to stability-oriented business customers.

"There's still a lot of testing going on—a lot of back and forth going on between our software partners and us," says AMD's Danny Shapiro. "Our current generation of hardware is of course Vista-ready, and Q2 is our target for rolling out the certified drivers that are already available for other operating systems."

And even though NVIDIA is in the middle of launching an exciting new line of DirectX 10 workstation cards with unified shaders and unprecedented performance, its newest ForceWare driver lacks support for SDI or Genlock. Several professional rendering apps are specifically listed as not supported in Windows Vista. And while the entire Quadro lineup is listed under driver support, there's still no mention of software certification.

Foreboding as all of that might sound, ATI and NVIDIA have clearly put a lot of work into software development for the new operating system. Days after Vista emerged, I was able to drop an NVIDIA Quadro NVS 440 into my workstation, install the operating system in record time, and boot to a gorgeous display. The RTM driver posted on NVIDIA's site had some clear stability issues, but the subsequent 100.65 release cleared those up. The Aero glass interface runs smoothly, even on the NVS 440's 128MB frame buffer—the bare minimum requirement. Professional graphics under Vista is emerging, but remember that it's a work in progress.