Computex 2006 Asymmetry and demos
By Charlie Demerjian in Taipei: Tuesday 06 June 2006, 08:49
TO THE SURPRISE of absolutely no one, ATI launched some fancy physics on its GPU technology today. Now we know what that third slot they ordered mobo makers to produce is for. Marketing terminology aside, they are supporting Havok with the Havok FX physics on the GPU API. If the numbers they showed today hold up to public scrutiny, ATI may be a day late, but they don't look to be a dollar short.
Under the banner of 'Boundless Gaming' with a fiddler crab pining for Ruby as a mascot, ATI has some compelling points to make with today's launch. First, it was an Intel/ATI launch, and the focus was Conroe and X1900, but there will be an AMD version coming soon enough. With the games that utilize it 6+ months away, it is not a big deal, just PR. ATI showed off asymmetric cards running physics on both Intel and ATI chipsets.
In press briefings, you generally expect a best case to be shown by the presenter, not untruths, but selective numbers. In this case, ATI showed off numbers where the a lowly RV530 absolutely spanked a G70 and G71 for physics. When you look at the performance relative to a R520 or R580, it gets even more abusive. The numbers ranged from a worst case where RV530 was 'only' 4x faster than G70 to a best where R580 spanked the G70 by 15x. They also went on to show the X1600XT as 2x faster than Ageia, and the X1900XTX being 9x faster. Whatever the numbers, they can put up a demo with a lot of rocks rolling down a hill.
Big grain of salt time, there is no game that runs across Havok and PhysX right now, so any numbers appear to be theoretical. Same with games that run physics across NV and ATI hardware. Because there are enough asterisks to fill a phone book, I will reserve judgment until I can get my hands on all three, and there are hard, real-world numbers out there. It looks good, but......
Where ATI does have a clear lead is the asymmetry. You could run a game with 3 X1900XTXs, and I am sure ATI stockholders would appreciate that, as would power companies, but you don't have to. You can plug in anything from an X1600 series GPU on up, mix and match. If you have a trust fund, go for the gold, otherwise, pick what makes sense for you. X1300 CPUs will work, but by the time the games come out, X1600 will be the bottom of the barrel, so the X1300 line is not really worth bothering with.
So, lets assume ATI does indeed thump NV on the same API. Why? The slides that ATI showed credits the branch prediction hardware. If an instruction stream has a branch, ATI can process it with no overhead. The resulting op after the branch takes whatever it takes. The 'green' architecture as they call it is shown to have a 6 clock overhead. Any instruction has it's latency plus 6, leading to a lot slower execution.
Whatever the case, ATI is in the game again, and is in it long before anything that needs GPU physics is. When the software catches up, it will be game on for real. At that point, and only then, will we see who really is the big dog, until then, it is all slideware. µ
