"The number of polygons it takes to render the real world was underestimated," he says. Back then, people were figuring about 1 billion: 10 million polygons per frame, 100 frames per second. "We can give a silhouette like Angelina 1 billion polygons a second. The trouble is, there's a part of the human brain dedicated to recognising human faces and characteristics, so you need some extraordinary high-quality pixels at the back end of that."
He does a rapid calculation: average displays now are "a factor of 25 from where we need to be", and you need those pixels to be a factor of 10 or even 100 better than they are now. Total: 25,000 times better than now. And computing progress is slowing down But not graphics power, which Huddy says is still tripling every year. Grab a calculator, wave a few exponents, and you get his 9 years and three months.
Ten years ago, you were running Windows, Word, Web, and email. In 1992, Huddy was overawed by a $1 million Silicon Graphics Reality Engine; in 2002 it was outclassed by a $15 graphics card. In 1997 IBM's Deep Blue beat chess grandmaster Garry Kasparov with a teraflop of computing power; ATI's latest system has a teraflop on two cards. CPU manufacturers are constrained by size and heat; graphics chips can keep getting bigger for another few generations and remain a bit slower. "You can get to a hundred parallel execution units on graphics chips where you struggle to find more than two or three on a CPU."
But Angelina is more complicated than more pixels and finer rendering of skin tones if you want her to look real to a human brain whose facial recognition capabilities are so good it can distinguish between identical twins. Huddy recalls the animator from the movie Shrek whose Princess Fiona didn't look real enough until he put in the tear ducts at the corners of her eyes. "It made all the difference between a thinking person and a doll."
Angelina needs tear ducts - and also the colour change when your ears are backlit and light bleeds through. Texture-mapping, the early approach, doesn't model the tone of skin because light penetrates a little bit and comes back out, changing constantly in response to lighting, action, age, oiliness, musculature. "The mathematics behind the representation has to change in quite subtle ways to capture it. There is a great deal more to do than the obvious stuff."
When you look around the Net for Huddy's work you find that in 2003 he joined ATI from its chief competitor, Nvidia, saying frequently that ATI was more like the company Nvidia used to be.
"We're a more grown-up version of that these days," he says. In part what he meant is that he believes Nvidia has sacrificed a percentage of quality in favour of benchmark speed. "I'm a technical guy." He wants people to be blown away by the power in the graphics chips. "ATI is true to that route. We try really hard to build hardware that fulfills the promise of game computing and cinematic rendering." We're standing at the bottom left corner of a graph. We want to be at the top right. The curve, at 3x per year, rises exponentially - and then flattens. When it passes reality, it will be travelling very slowly. "The temptation will be to say we're there, just ignore that last bit. Concentrate on speed, not quality. But quality is the key to getting above that line." To Angelina. µ
Wendy M. Grossman's Web site has an extensive archive of her books, articles, and music, and a page linking to articles in this series. Readers are welcome to send email to email@example.com (but please turn off HTML).
Sign up for INQbot – a weekly roundup of the best from the INQ