This was quite a good read, so I thought I'd share.
The Scottish Highland claymore is a two-handed sword with a cross hilt, is about 55-inches long and weighs about 5.5lbs. The Katana, as used by Japanese Samurai, is 38â€“40 inches long, and weighs less than two pounds. Interestingly, both are two-handed swords. While both swords certainly require skill and training to use well, the claymore's advantage lie in power and reach, while the katana's strengths lie in speed and precision.
Those differences between the two types of swords are pretty good metaphors for what's happening in the GPU market today. The two major competitors, Nvidia and AMD's ATI division, have taken substantially different approaches to product design. For the first time in a long time, the choices between GPUs are starkly different.
On the one hand, we have Nvidia's new GeForce 280 GTX. Weighing in at 1.4 billion transistors and a gigabyte of video memory, the 280 GTX demonstrates what you can do when you throw massive transistor budget at the 3D graphics problem. Of course, Nvidia will talk up the other capabilities of the 280 GTX, including CUDA and PhysX. We'll get to those shortly.
On the other hand, we have perpetual underdog AMD. Their latest GPU, the Radeon 4850, performs quite well, and is priced right, too, at under $200. Even at $200, it's likely that there's some price headroom, as the 4850 die size is considerably smaller than that of Nvidia's 280 GTX.
Good price and performance is certainly important, but the ATI team's approach to solving the problems was more interesting. Rather than throw massive transistor counts at the problem, they were given specific cost and die size targetsâ€”in addition to a fairly aggressive minimum power envelope.
At the same time, the performance target was pretty aggressive, including substantial improvements in anti-aliasing performance. Based on our initial benchmark testing, ATI delivered in spades. Is it as fast as a GTX 280? Nope, though it approaches the GTX 260 in some tests. And when you pump up AA, performance doesn't fall off a cliff.
This last bit is pretty important. A gamer who pays $200 for a video card is unlikely to spend over $1,000 for a monitor. It's more likely they'll have a sub-$400, 22-inch display with 1680x1050. At those resolutions, turning up AA is more useful than a few more frames per secondâ€”provided performance stays above that magical 60 fps number.
The 4850's approach to power is also interesting. Aggressive clock gating is used throughout the chip, which allows a card with the performance of the 4850 to get by with just a single PCI Express power connector. While Nvidia has dropped the price of the 9800 GTX to the $200 level, that card still requires two power connectors, and is a double-wide card.
ATI isn't ignoring the non-graphics market, either. At their editor's day event, they took pains to show off general purpose applications, right down to the same next-gen Photoshop demos and a video transcoding app (albeit different than the one Nvidia showed.) And while Nvidia can currently claim leadership in Folding@home, the app has yet to be fully tweaked for the Radeon 4000 series.
This may sound eerily similar to the marketing fight going on in CPUs between AMD and Intel. But it's really not the same thing. AMD's CPU guys have been doing a pretty good job of making lemonades out of lemons, including today's launch of the 2GHz Phenom 9350, a 65W quad core CPUâ€”a power envelope Intel's 45nm quad-core Penryns have yet to match.
Listening to Scott Hartog, the chief architect for the Radeon 4000 series talk, it's clear that they made a very conscious decision to approach the problem in a more elegant way. The end result is a highly efficient GPU. Nvidia may claim "better performance per shader core," but ATI's got efficiency where it countsâ€”better performance per watt.
The feature war goes back and forth, too. Nvidia: We've got PhysX! ATI: We've got hardware tessellation! Nvidia: We've got CUDA! ATI: We've got DirectX 10.1. And so it goes.
So who's going to win the war? We opened with a comparison of the highland claymore versus the katana. Both are capable weapons, in the hands of the right warrior. What it will really come down to over the long haul is how well each company executes on its vision. Nvidia has greater resources, but ATI has proven itself nimble and focused.
And don't forget, there's a third weapon that will enter the fight in about eighteen months: Intel's Larrabee. Will Larrabee be a more advanced sword, or a different kind of weapon altogether? I'm sure both Nvidia and ATI wonder about that, too.