First Look: AMD's ATI HD 4850 GPU

Think of the ongoing battle between nVidia and AMD as one painfully long prize fight to see who can deliver the better graphics processing unit. In the last round, nVidia delivered a series of 8800-based haymakers that AMD only now is recovering from. And while nVidia continues to swing hard with its GTX 200-series cards, the spunky ATI Radeon HD 4000 boards deliver a gut punch in the form of solid performance at lower price points.

The latest in the HD 4000 line--the Palit-branded, 512MB HD 4850--is currently in our Test Center, and is an elegantly designed, single-slot board that already looks to have top marks in the sub-$200 weight class.

AMD's plan: Create an easily scalable GPU that can stack up on a single card, so if you need more horsepower, you have an affordable upgrade path. With GPU upgrades, you can either buy multiple cards, go the route of high-powered dual-slot solutions--or take the hard-core gamer route and do both, blowing a small fortune in the process. AMD, though, is banking on mainstream users who hesitate to drop more than $300 on a discrete graphics board. So who is coming out on top? It's the classic standoff of brawn versus grace. Let's go to the tale of the tape.

The Battle Beyond the Graphics

Both GPU makers continue pursuing a card that does more than just paint a pretty picture. By releasing its CUDA SDK (software developers' kit), nVidia--as explained in our earlier story on the GTX 200 series--achieved a good head start in tasking the GPU with nongraphical tasks. That means it can promote this balanced computing model. The general idea: You shouldn't need to blow money on a top-notch CPU when you can partner a midlevel one with a good graphics card. The results in some tests right now are fascinating--encoding video files at a pace at least twice as fast, or just being able to manipulate photos (and 3D images) faster.

Don't count AMD out. The company can now pull off many of the same tasks with some of the same programs. One telling example: Adobe was on hand at both nVidia and AMD demos to show how its new version of Photoshop will work better thanks to GPU acceleration. While spokespeople are unable to say which graphics platform works better, the program at least provides one potential apples-to-apples test down the road. For now, the best shot for a fair comparison will be the full AMD-friendly client for the distributed-computing software called folding@home, when it's ready. That way, we can see how the software behaves on nVidia, AMD, and CPU-bound tests. Until then, it's a lopsided battle in nVidia's favor.

Then there's physics. nVidia may heavily tout the $400-plus GTX 200's built-in physics processors, but we're not yet seeing many developers using PhysX (that'll change soon, though). On the other end of the spectrum, AMD hedges its bets with partnerships. AMD has been talking about its GPU doing more than just graphics for ages, and recently the company announced that it is working with Havok's physics engine. Is it just me, or is it interesting that the Intel-acquired physics software maker is now in bed with the other big CPU maker as well? Combine that with the fact that nVidia makes a strong case for not buying top-end CPUs and I'm seeing a potential royal rumble brewing: Do both Intel and AMD see nVidia as a common foe? But I digress.

The Numbers Game

nVidia's G92-based 9800 GTX props up specs like a 675-MHz core clock and 2.2-GHz DDR memory strapped into a dual-slot card that requires two six-pin power connectors and a team of horses to run. By comparison, AMD's new board, built with a 55-nanometer manufacturing process, is all about efficiencies. It's got a lower core clock (625 MHz) and a lower memory clock (2 GHz) and yet manages to hold its own. That's due in no small part to the 800 stream processors (as compared to the 9800 GTX's 128). Not bad for a single-slot card that requires only one six-pin power connector.

Now, the caveats from last time still apply: We are continuing an initial shakedown run of a new graphical gauntlet. We are still hand-picking titles--current and upcoming--that'll challenge new hot-rod rigs, so that means no official scores on these graphics cards just yet. (Speaking of which, if there are games you'd like to see submitted into the official test list, let us know. Put your thoughts in the comments field!) All right, enough jibber-jabber. Touch gloves and come out fighting.

In Crysis, the Palit HD 4850 pretty much matches stride for stride with AMD's pricey--and beefy--last-generation card, a Sapphire HD 3870 X2. In Unreal Tournament 3, though, it's a whole other story. At both 1920-by-1200 and 2560-by-1600 resolutions, the HD 3870 X2 pulls ahead by 12 and 19 frames per second, respectively.

How does it hold up against nVidia cards? The 8800 GT runs the same numbers in Crysis until you turn on antialiasing. That's when the HD 4850 starts pulling ahead by a handful of frames. Hang on a second. What's that commotion in nVidia's corner? Just as AMD rallies, nVidia sucker-punches by dropping the price on overclocked 9800 GTX boards by 100 bucks, bringing 'em right in line with the HD 4850.

Shrewd move! That's like having Mike Tyson lose a couple pounds and shoving him into a lower weight class. Just as luck would have it, we managed to get a board for testing. The results: Trading blows, the XFX-built GeForce 9800 GTX XXX and the HD 4850 end up in a dead heat. Both cost the same (as of 6/26/08), and both post roughly the same performance numbers. AMD scores a couple extra frames in Crysis, and nVidia takes it right back during Unreal Tournament 3.

Worried about running up the electric bill during this slugfest? During all these runs, our test bed fluctuated between 156 watts while idle to 241 watts under load with the HD 4850. The XXX broke a sweat keeping up with the HD 4850, running about 20 watts hotter both when idle and when playing Crysis.

AMD manages to score solid hits, but ultimately this contest is still too close to call. And so the fight goes yet another round.