Slashdot Mirror
nVidia's GeForce 256 Breaks Out; changes 3D world
Hai Nguyen writes " nVidia officially unveiled the GeForce 256 (the chip formerly known as NV10).
Its architecture emphasizes both triangle rate and fill rate, so the chip can render 3D landscape with highly detailed 3D environments and models, and smooth framerates. Go get the full info." Holy moses. I want one. Now.
Damn. I was hoping this would get linked up on the front page. :) Oh well, I took a trip to NVIDIA last week, and I'd love it if you guys checked my article out: riva extreme - geforce 256 coverage
- 15M triangles/sec - sustained DMA, transform/clip/light, setup, rasterize and render rate.
- 4 Pixels per clock (4 pixel pipelines).
- 480M pixels/sec fill rate - 32 texture samples per clock, full speed 8-tap anisotropic filtering.
- 8 hardware lights.
- 350 MHz RAMDAC.
- Most feature complete for DX7 and OGL - Tranform & Lighting, Cube environment mapping, projective textures, and texture compression.
- Will utilize 4x AGP performance with Fast Writes , which enables the CPU to send data directly to the GPU (1 GB/sec transfer rate), increasing overall performance and freeing the system memory bus for other functions.
- 256 bit rendering engine.
- Highest quality HDTV (High Definition Television) video playback.
- High Precision HDTV video overlay.
- 5 horizontal, 3 vertical taps.
- 8:1 up/down scaling.
- Independent hue, saturation and brightness controls in hardware.
- High bandwidth HDTV class video I/O.
- 16 bit video port.
- Full host port.
- Dedicated DMA video.
- Powerful HDTV motion compensation.
- Full frame rate DVD to 1080i resolution.
- Full precision subpixel accuracy to 1/16 pixel.
Snipped from www.bluesnews.comSo basically nVidia chose to make a high fill-rate card with hardware lighting and transforms (geometry acceleration). These aren't innovative directions -- they were the obvious ones. None the less, the other major player, 3dfx, has pulled back from these choices. I'll explain why:
... everything is (we once demonstrated that it's physically impossible under DX6 to be faster than a Voodoo3 under Glide). There are some downsides: if you want to do crazy weird stuff with your lighting (eg. wrong faster stuff, funky effects) you may not be able to get it to work. Similarly with geometry -- special fast cases will become normal cases. So there may be a 50%-100% gain in triangle rate, but it's unlikely geometry acceleration will ever be able to provide much more than that.
... 3dfx talked about putting a geometry accelerator on V4 but I believe they backed off from it. Voodoo4 is however still an SST and therefore still a true descendent of the original Voodoo chipset conceived as a flexible, long-term solution for both PCs and arcade games.
... it may be finally time to kill some very old paradigms in 3d hardware...
nVidia has a card which can do supported operations fast. It obviously has a lot of fill. It'll be a good board. Of course it'll still be slow in D3D
nVidia seems to have chosen not to support the hardware bump mapping of the Matrox G400, an extremely high fill (runs beautifully bump mapped in a window in 1600x1200x32bpp) card without geom accel. 3DLabs' long awaited Permidia3 will also have some kind of hardware bump. IMHO this is a relatively flexible feature -- you could do a lot with it. It remains to be seen how flexible nVidia's lighting and geom turn out to be.
I'll be impressed if D3D ever delivers real hardware geometry benafits. We have yet to see a single benefit of DX6 over DX5 (not screwing with the fp control word especially) actually work. I'm highly suspect of anything MS sez.
So what about the remaining behemoth, 3dfx? Their Voodoo4 is supposed to be an extremely high fill card (fill has always been their hallmark). It may not support any more hardware features (eg. bump, lighting, geom accel), but it will fill like crazy. It's supposed to do full screen anti-aliasing
I'm eagerly awaiting the new generation. But I expect the real crazy stuff to start happening in the following generation
3dfx is rolling out another chip, as people have been talking about for a while. It is rumoured to be at 0.22 micron too, and will have hardware geometry processing.
S3 already rolled out a new chip - at 0.18 micron. It too has four texel engines and hardware geometry processing.
IMO, the S3 chip is actually the one to worry about. Architecture may or may not be great, but at 0.18 micron it may outperform nVidia and 3dfx's offerings just on linewidth. ATI did something similar when it rolled out the Rage 128, if you recall.
What I'm waiting for is the release of the GeForce or (insert name of 3dfx's offering here) at 0.18 micron. However, I'll probably be waiting a while.
That's easy- I presume you're rotating the tree realtime? _All_ that requires is that the tree can be cached on the card, which is then issued commands. ;) ) we have a comparable problem- there are 4X the voodoo cards as anything else, because of availability, and we're getting 'em off you PCers who are buying nVidias ( ;) dirt cheap, too! ), but Apple only supports ATI- so many important development tools are _not_ supporting 3dfx or Glide, and we are once again suffering the recurrent apple disease of Thou Shalt Use Only One Solution- in this case, ATI 3d acceleration. And I personally like 3dfx rendering better than even _TNT_, but this helps me not. (reading User Friendly I have been!). ;P
Why only one tree? What program, exactly, did this? There are some very serious questions to ask about demos like this. I, too, write software and try to come up with impressive claims. I can legitimately say that I'm writing a game with a ten million star universe with approximately sixteen million planets, of which the terrestrial ones (hundreds of actually landable-on planets) have terrains the size of the earth at 3 dots per inch for height information.
This is misleading as I'm doing it _all_ algorithmically- it's fair to ask 'well, what does it work like?' but nonsensical to imagine that somehow I'm messing with kajilliobytes of data. It's faked. (I have stellar distribution whipped, am working currently on deriving star types, slightly modified according to actual galaxy distributions- main current task is to come up with RGB values for the actual colors of star types, as this is more like white point color temperatures than anything else- very close to updating my reference pictures.
At any rate, will you believe me when I say that this reeks of demo? It wouldn't be that surprising if they used _all_ the capacity of the card to do that one tree. _I_ would. Might that be why there is only one tree and _no_ other detail at all (one ground poly, one horizon)?
More relevantly, what was used in doing that? If it was vanilla OpenGL, then okay, I concede this is very big. If they had to write their own software to do that, then you have a problem. Here in Mac land (also LinuxPPC land
You guys are looking at exactly the same situation here. Be damned careful. If you go with a proprietary technology you will fragment, and your developers will be faced with tough choices and could end up writing nVidia-only much as some developers in Mac land are writing ATI-only. This is bad. Do I have to explain why this is bad?
Let's get some more information about exactly how you operate this geometry stuff before getting all giddy and flushed about it, shall we? I don't see how software will use it without rewriting the software. And when you do that- it's an open invitation for nVidia to make the thing completely proprietary and lock out other vendors.
Or maybe they'd give the information out to people at no cost and not enforce their (presumed) patents for a while, only to turn around a year from now when they've locked in the market, and start bleeding people with basically total freedom to manipulate things any way they choose? But of course nobody (GIF) would think (GIF) of ever doing (GIF!) a thing like (GIFFF!) _that_...