ATI put GDDR4 on the Radeon X1950 XTX last month and reclaimed the single-card performance crown, which normally would be the story. It is not the story. The story is that both of these companies are about to throw out the architecture they have spent five years refining.
Everything shipping right now splits the work into separate pixel shader and vertex shader units, fixed in hardware at design time. That was a reasonable decision when nobody knew what the ratio between the two should be. It is now a liability, because the ratio changes constantly depending on what the scene is doing, and a card with idle vertex units while the pixel units choke is a card wasting silicon you paid for. Unified shaders fix that by pooling the hardware and assigning work dynamically.
The interesting detail is that this has already shipped, and not on a PC. The graphics part inside the Xbox 360 is an ATI design with a unified shader array, and it has been running in living rooms since last November. Console hardware is usually a generation behind the desktop. This time it was a preview, and a fairly public one.
Both vendors have parts coming that make the same jump on the desktop, and they are arriving alongside DirectX 10, which requires it. That is the part which complicates the buying decision for anybody reading a benchmark chart this month. A card bought today runs today’s games well and will not run DirectX 10 features at all, ever, because the capability is not there to enable in a driver update. Whether that matters depends entirely on how quickly developers move, and developers cannot move until the API ships, and the API is tied to Windows Vista, which has already slipped past its own launch window more than once.
My read is that the DirectX 10 argument is being oversold to people about to spend money. The install base for Vista on gaming machines will be small for most of next year. Studios building for a mass market are going to keep targeting DirectX 9 for a long time because that is where the customers are, and the first wave of DirectX 10 titles will mostly be existing engines with a lighting mode bolted on. If you need a card now, buy for the games you own. If you can wait until the new architectures land and the drivers stop being embarrassing, wait.
Underneath the product cycle there is a much larger thing happening, which is that AMD is buying ATI. That deal reframes the entire competitive picture. Nvidia has spent years as an independent graphics company selling into platforms other people control, and it is about to be facing a competitor that owns the CPU, the chipset and the GPU and can price all three as a bundle. Intel, meanwhile, ships more graphics parts than either of them by volume, all of them integrated, all of them terrible for anything with a frame rate, and Intel has never seemed especially bothered about that.
If I were Nvidia I would be less worried about the X1950 than about the possibility that graphics stops being a card you buy and becomes a feature of a platform you already bought. That is a slow threat rather than a next-quarter one, and the enthusiast market will always want a discrete card, but the enthusiast market is not where the volume is.
The other thing nobody wants to talk about is power. Look at what these cards now demand from a supply, look at the dual-slot coolers, listen to one under load. We have quietly reached a point where a high-end graphics card dictates the case, the power supply and the noise floor of the entire machine, and the industry response so far has been to ship a bigger fan. Performance per watt gets discussed in server rooms and ignored completely in gaming, and I do not think that holds for many more generations.
Buying advice, such as it is: the current top-end cards are excellent and they are also the end of a line. That is not a reason to avoid them. It is a reason to buy the one that is good value rather than the one that is fastest, because the fastest card of a dying architecture is the worst possible use of money. I will be curious to see how the first unified parts actually perform rather than how they benchmark, and I suspect the answer arrives later than anybody currently promising it would like.