Jeg fandt lige lige ekstra NV20 (GeForce3) info på Aces Hardwares forum. Dette er skrevet af "General Lee D. Mented" (fra http://www.aceshardware.com/board/general/read.php?message_id=20007786)
There's been several details leaked about NV20 (Nvidia's next chip, I don't think it's officially Geforce 3. The only other name I've heard is "Toy Story on a Chip"). It may be rumors but rumors aren't always unfounded.
Poly rate: 150Mpoly/sec
Explanation: a "poly" means triangle usually since you want this number to be high. Technically you can make a triangle with 2 new vertices if you have at least one vertex already in existance by connecting to the existing geometry. Hence, if NV20 has a 300mhz clockspeed (probable), and transforms 1 vertex per cycle (probable), 150Mpoly/sec. How to get all those polys over AGP? You don't, this chip is also rumored to have hardware curved surface support, hence most of these polys will be constructed locally from high level geometry. Note this also matches the Xbox specs floating around.
Pixel rate: 1200Mpixel/sec
Explanation: Obviously Nvidia wants their next product to beat their current product in every possible way (Geforce2 ultra). This means exceeding 1000Mpixel/sec. Also 300mhz x 4 pipelines = 1200Mpixel/sec. This reinforces the 300mhz expected clockspeed.
Textel rate: 4800Mtextel/sec
Possible: yes but with a hefty price tag
Probable: unknown given current memory situation
Explanation: Nvidia is rumored to be moving to FOUR textures per unit on this chipset (gee maybe they'll dig up my name entry for GeForce and call it Nvidia Quad Damage), as they anticipate newer engines (read Doom3) will use more texture passes (and according to Carmack they're right). 4 textures per pixel implies 4x 1200Mpixel/sec = 4800Mtextel/sec. Can they deliver this? I have no idea. It'd take some massive memory subsystem work or a bigass onboard cache.
Memory bandwidth: 19.2GB/sec ?
Possible: yes, but expensive as hell!
Probable: not unless they want 4 digits in the price or have an ace in the hole
Explanation: 4800Mtextel/sec 32bits per textel implies 19.2GB/sec. This would take a 256bit bus running with 300mhz DDR (600mhz) effective to deliver it. Or a 512bit bus at 150mhz DDR, which is the only currently feasible system I could think of, outside of embedded memory. One would make the pin count huge, driving up board cost. The other would make the die size huge, driving up chip cost. I can't see a way around this problem given current technology. However Nvidia is fairly good about delivering promises.
Tile rendering? Maybe. It's been hinted that this upcoming chip will have "zero or near zero overdraw". That usually means a tile architechure but maybe there's other ways to do it via more agressive poly culling.
Onboard compression? Putting compression in the actual pixel rendering stage would buy alot of memory bandwidth outside the chip. Possibly the 4:1 change from the current hardware, but only with compressed textures. Given how crappy S3TC is I pray this isn't the case or they use a different algorithm. It's a possible solution as it cuts memory demand back to 4.8GB/sec for 4:1 textures.
Shaders. Definately more shaders. And better support. Vertex shaders will be added (see DX8 previews), which will allow really neat stuff like say damage geometry instead of just damage skins, that's generated from force calculated from impacts. Or really nice flowing water and fabrics.
box is still for Fall 2001, and is supposedly going to use "an NV25 graphics core capable of 150Mpoly/sec". This will probably not be the expensive highend shipping product at the time. It will probably be a scaled down version of the revised core on a newer process. Think Geforce vs Geforce2 MX. What's the MX's specs like compared to the origional? What do you think highend NV20 at shipping this winter vs budget NV25 next fall will be like? =)
Also developers need a platform to start testing games on if Xbox wants to launch with any titles. This requires 2 things: 1. DirectX 8, which is due to ship soon I believe, and 2. Hardware capable of approximating the performance and behavior of the Xbox engine. See NV20. NV20 will probably also be the only card to be able to run DX8's new TrueHAL mode where everything is direct to hardware. This means hardware vertex and pixel shaders as well as transform and light. It really needs to ship about 9-12 months ahead of the system launch for a proper development cycle to occur, which would be by the end of the year if Microsoft is still on a september target. =)
I say expect a simultaneous NV20 and DirectX 8 announcement (and IMMEDIATE chip availability in volume, i.e. it's been stockpiled thanks to TSMC doing well on the fab side) sometime this month. This would give a fairly short OEM ramp time (less if OEMs are currently designing under NDA) of 4-6 weeks, which could still put us on either side of thanksgiving weekend (November 23rd to 26th for you non-US readers) to see cards on shelves.