First Look: NVIDIA GeForce 3 (2/6)

The GeForce 3's specifications are really anything but spectacular... 200MHz core speed (GF2 Ultra=250MHz) and a memory speed identical to that of the GeForce 2 Ultra (460 MHz), even though the 3,8ns memory chips are specified to run at speeds up to 500MHz. The amount of memory also remains unchanged at 64MB - nothing exciting nowadays. Amazingly, the new chip has more than twice the number of transistors of its predecessor, a whopping 57 million (GF2=25 million) and is manufactured on a 0.15ĩ process. The picture at the bottom of the image on the left shows the size of the die in relation to a 10 Pfennig coin (which is roughly the size of a nickel, for our American readers :-) ).

Despite its nominally lower clockspeed, the GeForce 3 will (or letīs say should) be noticeably faster than the GeForce 2 Ultra, thanks to a couple of tricks that NVIDIA has pulled out of its hat. Most of these attack one problem specifically that has become most video cards' primary bottleneck: Memory bandwidth and AGP transfer rate. NVIDIA uses three of these "tricks" together and calls them "Lightspeed Memory Architecture":

• Crossbar Memory Controller: The internal 256 Bit memory bus is divided into four paths of 64Bit each. This way, smaller data packets don't automatically block the entire bus. When larger data sets need to be transferred, these subdivisions can be reconnected, giving you a 256Bit wide bus. Ideally, the CMC is supposed to be 4x faster. NVIDIA says that the 64-bit data paths are sufficient for about 75% of all memory transfers that occur.

• Lossless Z-Compression: One of the biggest memory bandwidth hogs is Z-Buffer data, since it is read from memory and written back to it for every rendering cycle. The GeForce 3 transfers this data only after it has been compressed at a ratio of 4:1 (lossless).The compression and decompression takes place completely in hardware and thereofre reduces the strain on the meory. While the feature is strongly reminiscent of ATi's Hyper-Z, NVIDIA maintains that their technique is something completely different...

• Z-Occlusion Culling: During the last months we have witnessed rumors of (leaked) drivers supposedly supporting HSR (Hidden Surface Removal) on GeForce 2 cards flying around the web. That's all it was though - a rumour. So what is HSR, you ask? Traditional (so called "brute-force") graphics chips render a lot more polygons than you'll ever see in the completed scene - objects behind a wall, for instance. (Victims of the legendary ASUS cheat/see-through drivers will know what I mean). Z-Occlusion culling is supposed to help the GeForce 3 predetermine the visibility of any given texel. If the texel is judged to be invisible, it won't be rendered and thus frees up space in the framebuffer as well as saving the chip a read/write to/from the memory.

According to NVIDIA, the LMA features alone will make the GeForce 3 up to 4x faster than a GF2 Ultra. If these claims are borne out in real-world scenarios remains to be seen.

Another major change involves anti aliasing. This feature is supposed to reduce the jagged edges you see on most 3D objects in games (regardless of resolution, although they become less noticeabnle at a higher res.) Originally, this feature was pioneered by 3dfx, and was to become the Voodoo 5's main selling point. This sparked a discussion about the necessity of FSAA, dividing gamers into two very distinct factions. On the one side are those who prefer to play at higher resolutions, where jagged edges become less and less noticeable. On the other side are those that decided to drop the resolution down a level so they could use FSAA at acceptable framerates. (Another deciding factor is also monitor size - if your moonitor can't display 1600x1200, you don't really have all that many options) GeForce 2 boards also support FSAA, but it carries a massive performance hit in the implementation used. Suddenly, with the GeForce 3, FSAA has become one of NVIDIA's favorite features.

Instead of using SuperSampling (in a nutshell, this means scaling an image to a higher resolution and then downsampling it back to the original one), the GeForce 3 employs a different mehtod, called MultiSampling. While the difference is in the details of the implementations, MultiSampling has a very tangible advantage: since it only needs to read a texture once, it cuts down on a lot of memory traffic. We'll take a look at the details including a comparison of the two techniques in a later article.

A nice side effect is that these techniques will also speed up current games, not just DX8 games.

Copyright: 27.02.2001 -   RIVA Station 2001 - Lars Weinand
No Copy without Permission!

Translation by Benjamin Kraft

URL of this Article: www.rivastation.com/gf3_e.htm - If you want to link to it, please use this URL! :-)