For a while there, the graphics market was looking like a one-vendor show. NVIDIA unequivocally dominated the high end with its GeForce 8800-series cards. AMD’s Radeon HD 2900-series cards just couldn’t keep up, and the subsequent 3800s didn’t fare much better. As NVIDIA’s first-generation unified shader architecture aged, it maintained a performance advantage that AMD just couldn’t contest. Instead, AMD focused on scaling back power and reducing the die size of its GPUs. The company worked to optimize the value of its graphics products. But boasting big-time value in graphics is like being known for having a great personality in Hollywood—it won’t get you many magazine covers.

Despite the fact that AMD’s former desktop flagship, the Radeon HD 3870, wasn’t a major head-turner, it did manage to win over customers in the $200 price range. Meanwhile, NVIDIA was left to reign over the high-end market, where its GeForce 8800 GT, GTS, and GTX cards let enthusiasts granularly choose the fastest graphics card for their money. For more than a year the situation remained the same. AMD quietly composed its mainstream platform message using the Phenom X4, 790FX chipset, and Radeon HD 3870 as NVIDIA swung for the fences, hitting a couple of home runs of its own in the GeForce 8800 GT and Hybrid SLI technology.

The Rise of a Contender

Of course, it’s easy to get comfortable when there’s nobody ahead of you and your nearest competitor really isn’t very near at all. Perhaps that’s why NVIDIA’s lineup evolved slowly through 2008. Whatever the reason, today’s reality is quite different. Though AMD isn’t gunning to sell the fastest individual GPUs, its new RV770 processor is, without a doubt, a massive step up. In fact, your customers can now get faster frame rates at lower prices with AMD’s technology than the old incumbent—to a point.

THE MAKINGS OF PERFORMANCE: Inside that smooth piece of silicon you’ll find 800 stream processors—enough to more than double the speed of last-generation’s graphics cards..

Though the new Radeon HD 4850 and 4870 cards are earning AMD so much attention now, it’s actually the RV770 graphics processor powering them that deserves most of the respect. A product of immense optimization and cooperation from six international offices, RV770 was a particularly challenging design because AMD had to build it without the benefit of a manufacturing process shrink. It used the same 55nm node already employed by last-generation’s Radeon HD 3870. The fact that RV770 serves up so much more performance is a real wonder.

Because RV770 centers on a number of architectural elements derived from the R600-family, some of its internals look familiar. The GPU offers way more computational horsepower, though. For instance, RV770 sports 10 SIMD cores, each with 16 stream processor units. You’ll find five ALUs per unit, bringing the chip’s total to 800. In comparison, RV670 comes with 320 stream processing units. The increase is enough to boost RV770’s theoretical performance up to 1.2 teraflops.

AMD also equips RV770 with significantly more texturing capability, and it quantifies the increase by claiming the redesigned texture units crank out 70% more performance per square millimeter of die than the chip’s predecessor. Surprisingly, AMD’s flagship GPU is fed by a lightweight 256-bit memory bus divided into four 64-bit interfaces, each tied to its own render back-end. We’ll talk more about the secret weapon that AMD uses to maximize RV770’s bus as we dissect the different models available at launch.

The GPU itself is incredibly large, though not nearly as big as NVIDIA’s latest chip. As mentioned, it’s manufactured at 55nm and now includes somewhere in the neighborhood of 965 million transistors. Naturally, AMD balances the big processor’s thermal characteristics using on-chip power management technologies, variable clock speeds, and a couple of different cooling solutions, which help define the two models now available.


Productizing RV770

Of course, resellers can’t sell graphics processors. You need cards built on those GPUs that combine the right memory configuration, physical dimensions, power consumption, and price. Right out of the gate AMD is offering two boards, the $199 Radeon HD 4850 and the $299 Radeon HD 4870. Relatively speaking, both cards are fairly mid-range. They’re going to appeal to very different customers, though.

The lower-priced 4850 is a more elegant solution. It populates a single upgrade slot with its slim heatsink and fan. Although the 4850 runs at a lower clock than the Radeon HD 4870, its 675 MHz core frequency is enough to serve up that teraflop performance figure. Moreover, 512MB of GDDR3 memory running at 993 MHz delivers enough bandwidth to keep the Radeon HD 4850 ahead of most of NVIDIA’s cards. And AMD’s clock targets are such that the board only requires a single six-pin auxiliary power connector to satisfy its 110W TDP.

Given its thermal envelope and form factor, the 4850 opens the door to some interesting options. For example, Shuttle sells an X48-based barebones system with two PCI Express 2.0 links, each wired for x16 operation. A couple of Radeon HD 4850s in a machine like that are going to yield an impressive gaming experience. Or, attach up to four monitors to the cards’ quartet of dual-link DVI ports, creating a powerful graphics workstation.

SLIM IS IN: The single-slot Radeon HD 4850 is great for cramped cases, where it still manages to crank out fast frame rates.

The Radeon HD 4870 is even more impressive. It leverages the exact same graphics processor at 750 MHz instead of 675 MHz. Incidentally, AMD claims there is plenty of headroom to push RV770 even further—a task your customers can safely undertake through the Overdrive panel in the Catalyst Control Center driver package. The higher clock speed necessitates a larger cooler, so the HD 4870 gets a dual-slot, blower-type implementation. Rather than circulating hot air like the 4850, AMD’s Radeon HD 4870 exhausts it to help keep ambient case temperatures to a minimum. And in order to cope with the faster clock speed and more powerful fan, the 4870 requires a second six-pin auxiliary power connector, meaning VARs will want to pay particular attention to the power supplies being bundled with the $299 card.

Like the Radeon HD 4850, the 4870 sports 512MB of onboard memory. However, instead of leveraging GDDR3, the flagship board uses GDDR5 memory technology to dramatically increase bandwidth. Though the GDDR5 ICs run at just 900 MHz, they push throughput up past 115 GB/s. Adopting GDDR5 was a notable gamble for AMD, but it looks like it’s paying off because the Radeon HD 4870 wields the 256-bit memory path of a mid-range board and the bandwidth of a true powerhouse.


NVIDIA Strikes Back

Actually, NVIDIA struck first, unveiling its GeForce GTX 260 and 280 cards just days before AMD’s own introduction. However, the fight is certainly on NVIDIA’s home turf—the high-end market it has dominated for so long..

The GTX 260 and 280 center on the same second-generation unified shader architecture, contained within a monstrous GPU referred to as GT200. Manufactured using TSMC’s 65nm node, the 1.4 billion-transistor chip is indeed massive. Fortunately, the cooling solutions enveloping both cards are designed to dissipate heat as quickly as possible.

And yet, GT200 isn’t a drastic departure from what came before, nor should it be. The GeForce 8800-series held NVIDIA’s commanding position for more than a year. Compared to the aging G80’s 128 stream processors, GT200 boasts 240. Don’t pit that number against AMD’s 800 stream processors because they mean different things. More significant is NVIDIA’s claim that the GeForce GTX 280 can do 933 gigaflops, slightly less than AMD’s Radeon HD 4870.

JOLLY GREEN GIANT: NVIDIA’s massive GT200 GPU, comprising 1.4 billion transistors, serves up the best performance available from a single board.

Even then, you can’t buy into theoretical numbers. After all, the GT200 has more texturing hardware and twice as many render back-ends (eight) attached to a memory bus two times wider (512 bits) than AMD’s RV770. The result, naturally, is a high-performance graphics processor worthy of succeeding NVIDIA’s last big GPU.

GT200 has a couple of tricks up its sleeve belying the massive coolers topping both new cards. To begin, although GT200 dissipates more than 200W of power under load, the thing idles at 25W. Compare that to G80, which lumbered along at more than 60W, according to NVIDIA. Because your customers will undoubtedly be idling more often than running under load, the GTX 260 and 280 should save energy compared to the older high-end boards. The secret is heavy clock gating, dynamic frequency adjustment, and power states similar to what you might expect from a Core 2 or Phenom processor.

On each of its new cards, NVIDIA complements the GT200 with a companion chip, which enables a pair of dual-link DVI outputs with support for 10-bit color per channel. HDCP support is, of course, included as well. NVIDIA is claiming HDMI compatibility through an adapter. However, none of the boxed cards we’ve seen come with it yet.


Selling NVIDIA’s GT200

Call the GT200 beefy. Call it big-boned. Call it healthy. The point is that it’s a large chip with some serious potential for heat dissipation—and that’s why the GeForce GTX 260 and 280 both sport large cooling solutions fully enveloping each card.

Technically speaking, the 260 is NVIDIA’s second-in-command, though it’s still priced in the high $300s and well-equipped to do battle with AMD’s best. In order to put a bit of space between the GeForce GTX 260 and 280, NVIDIA disables 48 of the GT200’s stream processors. The result is a GPU with 192 stream processors and a somewhat odd 448-bit memory bus, to which is attached 896MB of GDDR3 memory. Clock speeds on the board are also a bit wonky: 576 MHz for the core, stream processors running at 1,242 MHz, and 999 MHz memory.

King Of the hill: With its 240 stream processors and 1GB of GDDR3 memory, the pricey GeForce GTX 280 is tops for 3D performance.

The GeForce GTX 260 is a physically substantial card. It requires a dual-slot cooler, which means it isn’t necessarily your best option for small form factor enclosures. However, the board also manages thermals better than single-slot solutions by exhausting air instead of blowing it around inside your customer’s chassis. Interestingly, the entire board is enclosed in NVIDIA’s cooler—front, back, everything. The only pieces exposed are two six-pin auxiliary PCI Express power connectors and an SLI connector, which actually ships with a cap over it.

You’ll notice that the GeForce GTX 280 looks quite similar. Of course, its more potent graphics processor runs at faster speeds and consequently consumes more than 230 watts under load—enough to warrant one six-pin PCIe power connection and one eight-pin PCI Express 2.0 plug. The card’s 240 shader processors operate at 1,296 MHz, the core runs at 602 MHz, and the board’s gigabyte of GDDR3 memory dashes along at 1,107 MHz on a 512-bit bus.


The Cost to Be the Boss

NVIDIA launched the GeForce GTX 280 at a $649 price point; the 260 was originally supposed to cost $449, then $399. Clearly the company is under competitive pressure from AMD. And that’s good news for your cash-strapped enthusiast customers, who might like the idea of two high-end cards rendering cooperatively but not the thought of sinking more than a grand into graphics. Recent price drops have pushed the GeForce GTX 280 down closer to the $500 range while the GeForce GTX 260 is hovering around $330.

Though the pricing pressure might frustrate NVIDIA, which is likely paying a hefty premium on those 1.4 billion-transistor GT200 graphics processors, it was badly needed in order to balance a competitive high end. With AMD’s Radeon HD 4850 and Radeon HD 4870 cards costing $199 and $299 respectively, the GeForces were overpriced in comparison.


Multi-Card Madness

“But wait,” you say. “Aren’t the NVIDIA cards both faster than AMD’s more mainstream boards, built on a less complex architecture?”

The GeForce GTX 280 is indeed the fastest single card you can buy right now, though AMD does have well-known plans to usurp that dominance with a card featuring two RV770 chips onboard. Additionally, the battle between NVIDIA’s GeForce GTX 260 and AMD’s Radeon HD 4870 is a close one.

Where this all gets more interesting (not to mention more profitable)—and the reason resellers should be thankful we’re primarily covering cards in the $300 range here—is when you start looking at AMD’s and NVIDIA’s multi-card rendering technologies, naturally supported by both company’s latest architectures.

AMD’s CrossFireX supports configurations up to four cards deep. So long as you’re selling a compatible platform with the right number of PCI Express x16 slots, your customer can buy one card, say the $199 Radeon HD 4850, and then upgrade to a second as his performance demands increase. In a board with three or four PCIe slots, like Gigabyte’s GA-MA790FX-DQ6, third and fourth 4850s help scale performance even further.

The answer to CrossFireX is, of course, NVIDIA’s SLI technology. While SLI only scales to three cards, you could argue that a three-way GeForce GTX 280 platform might still be faster than four Radeon HD 4870s. But at the point where you start talking about $1,500 worth of graphics hardware, the real issue becomes finding games to tax such a powerful configuration. The takeaway is that customers now have access to as much rendering horsepower as they can afford from either vendor, making it almost pointless to argue over who has the fastest individual board.

Competition, How We Missed Thee

The introduction of AMD’s RV770 graphics processor is music to the channel’s ears. Not only are the cards centering on the chip priced attractively, but competing boards are quickly coming down in price in time for the back-to-school season. And it’s only a matter of time before we start seeing workstation-class boards based on the latest graphics architectures from both AMD and NVIDIA.