|
||
|
|
|
              |
|
 |
 |
||||||||||
|
|
|||||||||
By Chris Angelini |
||||||||||
Microsoft’s Windows Vista did an interesting thing to the desktop graphics market. By offering a 3D-accelerated desktop, the operating system presented AMD and NVIDIA with an opportunity to push higher-end hardware. The idea was that everyone, even businesses, would want the best Vista experience possible, and that meant taking a step up to more powerful graphics chips. So the two market leaders poured resources into their respective unified shader architectures. This wasn’t necessary to support Vista—-after all, you can drop a DirectX 9 board with 128MB of memory into a Vista box and still be Vista Premium-ready. But by incorporating all of the goodness wrapped up in those newer designs, an entirely new generation of DirectX 10 software based on Shader Model 4.0 is enabled for systems running under the only OS to support Microsoft’s latest API. Professional customers might be left scratching their heads, wondering why they would care about a wave of cards optimized for DirectX 10 when most of their software runs under OpenGL. The fact of the matter is they wouldn’t. Workstation folks running CAD software don’t pick up inexpensive desktop boards for the same reasons gamers don’t spend $2,000 on professional cards. One simply isn’t intended to do the other’s job. Although both leverage the same hardware, driver teams on each side of the fence create vastly different software packages. Clearing The Air Because AMD and NVIDIA were hurriedly putting the wraps on their unified shader designs (and because businesses are still approaching Vista with a very cautious eye), sales of discrete graphics cards tanked in the first quarter of 2007, according to Jon Peddie Research. The business in Q2 fared much better thanks to NVIDIA’s lower-cost GeForce 8000-series cards and AMD’s more recently launched Radeon HD 2000 boards. Apparently, everyone was waiting for those major transitions before making a move into discrete graphics. But it’s important to understand why the latest cards are faster and exactly where your customers can expect to see gains before diving into a new generation of professional boards based on completely reworked architectures. The performance boost that today’s unified shader architectures offer is a result of more efficient operation rather than running under Vista or utilizing DirectX 10. That’s an important clarification because business customers buying professional graphics probably won’t be making heavy use of DirectX. However, knowing that they’ll see a significant gain in the applications they use today just by adopting the new technology is a substantial value. There’s also the potential for extra speed as graphics vendors tweak developing driver code and software developers optimize applications to take advantage of the latest hardware capabilities that are not yet being used. So when a customer asks why he would want a new professional card based on an architecture most recently associated with Vista, DirectX 10, and gaming, the answer is simple. Tell him that according to industry professionals such as Danny Shapiro, senior marketing manager of workstation graphics at AMD, he can expect immediate performance increases in existing professional apps now and even more as the drivers continue improving.
Unified From Top To Bottom Seven months ago, we interviewed AMD to find out where it was taking its FireGL product lineup. NVIDIA had just launched the Quadro FX 4600 and FX 5600 ultra-high-end cards derived from the GeForce 8800 series and AMD’s own next-gen parts were still in the works. Representatives at AMD speculated that it would be at least six months before the professional market even started considering Vista. Because graphics workstations are not at all tolerant of instability, software drivers for the new operating system were spotty at best. ISVs hadn’t yet said much about Vista support, and the Vista transition seemed like something that would happen once the operating system’s quirks had been worked out. At the same time, NVIDIA was enjoying a lot of success with the Quadro FX 4600. The card’s allure has less to do with Vista or DirectX 10 and more to do with the pure efficiency of its unified shader architecture. Even with its $1,999 price tag, the card dishes out copious value in the way of screaming performance. Although it was a couple of months late with its newest generation of desktop cards, AMD is wasting no time at all translating a full family of graphics processors that employ unified shaders to the professional workspace. AMD is slipstreaming the momentum, already pushing unified designs along, and now has the first top-to-bottom lineup of FireGL products based on its fastest desktop graphics cores. The top end of AMD’s professional fleet is very similar to its flagship desktop model, with a handful of significant exceptions. To begin, the FireGL V8650 leverages 2GB of onboard memory-—more than any other card available. Dropped onto a 512-bit memory bus capable of delivering 128 GBps of bandwidth, you can imagine the design complexity that goes into the board. Of course, if we’ve learned anything from the past couple months of digging into memory and processors, a large graphics frame buffer means a potentially large hole in system memory while operating under the constraints of a 32-bit environment. You’ll want to install the board in 64-bit workstations with plenty of RAM to get around that. Like AMD’s Radeon HD 2900 XT, the V8650 centers on an R600 graphics processor, a second-generation unified shader architecture loaded with 320 shader units. Briefly, each shader unit is able to operate on either vertex or pixel data as the application and drivers see fit, serving as a more general-purpose processor than the distinct pixel and vertex shaders characteristic of older designs. Applications heavy on geometry or pixel information that would have previously left a lot of onboard resources idle can now take advantage of the whole chip’s power. AMD has the 720 million-transistor R600 manufactured on an 80nm process at TSMC. The FireGL V8650 consumes a lot of power, and the massive frame buffer probably doesn’t help much. All told, expect to feed the card more than 255W in addition to whatever your dual- or quad-core 2P workstations need. Also in the ultra high-end camp is AMD’s FireGL V8600, identical in every way to the V8650 except that it boasts 1GB of memory instead of 2GB. The smaller memory bank reduces the card’s appetite for energy, dipping it below the 225W mark. There’s also the $1,899 price tag, which puts the card well within reach of more budget-conscious workstation customers. Bandwidth, computational horsepower, and value-added features all remain the same.
The third card built on AMD’s R600 GPU, AMD’s FireGL V7600, sheds a little more memory in favor of a sub-$1,000 MSRP. Classified as a high-end board, the V7600 sports 320 stream processors but is fed by 512MB of memory on a 256-bit bus that more than halves memory bandwidth to 51 GBps. Dropping to 512MB and slowing the card’s clock speeds down has the positive side effect of reducing power consumption to less than 150W. Once you drop to the mid-range FireGL V5600, AMD’s R600 GPU is replaced by the RV630, an arguably more elegant 65nm chip with 120 stream processors. A smaller manufacturing process and less complex architecture have the combined effect of cutting the V5600’s power consumption down to below 75 watts. Priced at $599, the V5600 includes 512MB of memory on a 128-bit bus that moves a conservative 35 GBps. Even so, according to AMD’s own benchmarks, the FireGL V5600 is able to lay down more performance than previous-generation cards costing three times as much because of its unified architecture. The same graphics processor is even available to entry-level buyers at a $299 price point in AMD’s FireGL V3600. Sporting 256MB of memory on a 128-bit bus, your customer has a mere 16 GBps of throughput with which to work. However, the 120-processor GPU is still capable of doling out the speed. Best of all, less than 50W of power consumption means it’s possible to build entry-level graphics workstations in small form factor chassis. Value Across The Board There’s a tremendous amount of variation in performance from the inexpensive FireGL V3600 to the ultra-high-end FireGL V8650. However, AMD is doing a great job of preserving functionality from one card to the next. For example, both extremes of the family boast dual-link DVI outputs. You’d expect as much from a $2,800 board. But a lot of entry-level products round down to single-link DVI or VGA output in an effort to save money. Getting that high-end connectivity from top to bottom means you can sell any FireGL card with a pair of 24” 1920x1200 LCDs and expect the same crisp picture quality. All five of AMD’s new FireGL cards also offer HD component output, which attaches to any number of big screen displays. As you start getting into the chips powering each FireGL board, you’ll notice that the features remain constant. They’re all based on the same engine with stream processors yanked out here and clock speeds adjusted down there to make the best use of yields. However, the unified design that makes each card so efficient remains the same. A customer can buy one high-end CAD workstation with a V8650 and another box with the V5600 and expect consistent results from content shared between the two.
One component of that unification relies on hardware. But software plays an important role, too. AMD’s latest drivers address all of its FireGL cards and add an innovative feature called Auto Detect that simplifies switching between professional apps installed on one machine. Workstation cards typically include software profiles for specific titles. Those profiles flip switches and turn knobs within the software so when you run Softimage|XSI, the card is tuned for XSI. When you run 3ds Max, the drivers are best configured for that app. Auto Detect recognizes the most popular titles and automatically adjusts the drivers on the fly. Both AMD and NVIDIA already use similar tricks with their desktop cards in order to procure the most performance possible from them. However, Auto Detect is the first professional software feature able to dynamically optimize, even to the point of changing settings when multiple programs are running simultaneously. The Truth About Multicard Setups AMD’s CrossFire and NVIDIA’s SLI multicard rendering technologies are big hits on the desktop due to the extra performance they enable. When it comes time for professional rendering, however, both technologies prove a lot less effective. According to ATI’s Danny Shapiro, most workstation apps are run in a window, so the on-screen graphics workload is uneven. That naturally creates problems for two cards dividing the screen in half. Load balancing is only one of the issues, though. Apparently, the speed of a multicard setup can actually dip below the potential of a single card in some cases. No wonder AMD isn’t pushing cooperative rendering. The company is, on the other hand, advocating dual-card configurations. All of the new FireGL cards will run opposite another board in the same motherboard to enable quad-display outputs. NVIDIA’s boards also support tandem operation. But NVIDIA is still pushing SLI as a means of significantly improving performance. AMD contends that you should be pushing two cards for their connectivity benefits, not potential speed gains. AMD Gets Competitive Before it was owned by AMD, or even ATI, the FireGL brand stood for some of the fastest professional video accelerators available. Stiff competition from more regularly refreshed desktop architectures put the lineup under pressure, and it eventually fell behind.
When ATI acquired the brand, it began adapting its own desktop chips to the workstation market. The hardware evolved, the software drivers solidified, and performance scaled up with each new generation. The FireGL brand still hadn’t returned to its former glory, though. This time around, AMD is confident that its second unified shader design has the muscle at the top end, flexibility at the entry level, and dependability across the board to reclaim what was once a prestigious position for the FireGL brand. |
||||||||||
 Back to top |
||||||||||
Copyright © 2007 RAM Magazine. All rights reserved.
Do not duplicate or redistribute in any form. |
||||||||||
