Slashdot Mirror
ATI Radeon 9700 Dissected
Bender writes "The guys who laid out the future of real-time graphics a while back have now dissected ATI's Radeon 9700 chip. Their analysis breaks down performance into multiple components--fill rate, occlusion detection, pixel shaders, vertex shaders, antialiasing--and tests each one synthetically before moving on to the usual application tests like SPECviewperf and UT 2003. You can see exactly how this chip advances the state of the art in graphics, piece by piece. Interesting stuff."
The real question for those of us who don't run Windows is how well it works in X. What it ATI's attitude towards open source ? Are their specs public ? Do they provide drivers ? In short is there a reason to switch from nVidia when I upgrade ?
May contain traces of nut.
Made from the freshest electrons.
It's partly a working style issue. Texture-map people go out with cameras and photograph nice-looking surfaces, which they then apply to models. Or they paint the textures. Procedural shader people try to code up the "meaning" of a texture. Texture maps are created by artists; procedural shaders are created by programmers.
The basic problem with texture maps, of course, is that you can't get too close before the texture gets blurry and the illusion breaks down. In film work, you know where the camera is, so you can figure out how much texture detail you need. Games don't have that luxury; you can get close to a surface and blow the illusion.
Most film work other than Pixar's has used texture maps. There are exceptions, but they're typically for hair, fur, and water, where the problem is motion.
The price you pay for using procedural shaders is that they usually model surface, not detail. So you have to model the detail. Lots of it. Again, Pixar is notorious for this. ("We modelled the threads on the screws, even though you couldn't see them!")
Texture maps, bump maps, and displacement maps can be used to modulate procedural shaders, and that's probably how surface detail will be done, rather than getting carried away with building complex textures in some programming language.
Every NVIDIA card since the GeForce2 Ultra has had Linux drivers before they even hit the shelves. This is because Nvidia pay people to write and maintain the drivers. They might not have specs, but at least NVidia support your choice of operating system.
ATI release some specs, and that's all. They don't either bother writing drivers for their cards and they just hope someone else will - *maybe the weather channel, maybe soon, maybe later, maybe not for your specific card) or release binary-only drivers (great, at least they exist) that don't have anything like the performance of their Windows drivers. The UT2003 benchmark, if ran under Linux, won't even start on a Radeon 8500 (which ATI do have fast, binary only drivers for because its missing correct support for S3 texture compression. Which isn't exactly a new technology by any means.
So I can get Open Source 2D support for a Radeon 9700? Great. I'm sure 2D support is why people buy a Radeon 9700.
Vote with your dollars.
NVIDIA uses the same codebase for ther Windows, Mac OS ?, and Linux drivers. This same codebase will also be used for their FreeBSD drivers to come. Their unified driver architecture ensures that every platform the card runs on gets the latest version of the code and can take advantage of each card's features. So this is definitely a few notches above ATI who won't even produce drivers for my platform, let alone release full specifications to the public to write them.
As for the complaint that NVIDIA is no better than ATI because of a binary driver release: that is not NVIDIA's fault. NVIDIA tries to make as much of their driver open source as possible (which is kind of a necessity because of the plethora of kernel configurations out there). However, the closed-source portions are kept closed because of SGI's patents on OpenGL. Assign blame where blame is due, please.
Why bother.
Actually, it does matter with more FPS. Don't compare it to film, because even though they both use the term 'frame' they mean different things.
A 24 fps film means that each frame is recording 1/24th of a second. That means that if an object being filmed is moving fast enough, the frame will have motion blur. When strung together with the other frames, this will give the illusion of smooth movement. A 24 fps 3d engine, however, means that you have 24 static shots. There's no transition from point to point, unless you wind up rendering said inbetween shots. Or, put another way, a 5 fps film of a hand waving in front of the camera will produce five frames full of motion-blurred hand, which, when played, will look relatively smooth. A 5 FPS render, however, will have five static shots of a hand sitting motionless in space, and when played, the hand will appear to 'teleport' from spot to spot to spot.Or, put another way, record that hand with a standard camera shooting at five 'frames per second' not 'several frames, each 1/5th of a second exposure' and then string the negatives into a film reel, splicing in copies to make the whole thing last one second.
This is one of the reasons, I always thought, that 3dFX was trying to get their T-buffer out into the world, becuase then, yes, if you could LOCK the rendering at 30 FPS, and throw in motion and acceleration blur, it would still look better than a card rendering the exact same thing at 300 FPS.
Vintage computer games and RPG books available. Email me if you're interested.