GeForce-3 ULTRA Specifications

[RobRich]

I just received this info a few minutes ago. I do not have the latest nVidia internal roadmap documentation for beyond the GF-3 standard, so I can not verify the data. However, I usually do not post pre-release information unless the source is generally deemed reliable.

nVidia GeForce3 ULTRA

Fabrication Process: 0.15-micron
Transistors Count: 50+ million
Rendering Pipelines: 4
Texture Units per Pipeline: 3
Core Speed: 300MHz
Memory Speed: 500MHz (250MHz DDR)
Memory Bandwidth: 8GB/sec
Pixel Fill Rate: 1.2 giga-pixels/sec
Texel Fill Rate: 3.6 giga-texels/sec
Triangle Processing Rate: 150 million-triangles/sec

Catch ya' later,
Robert Richmond

[This message has been edited by RobRich (edited 03-07-2001).]

[Wilan Wong]

Seems to have very nice specifications, but I think the main bottleneck in that spec is definitely the 8GB/sec memory bandwidth. There is no way that the 3.6gtexel/s would be effective since it would hit the memory bandwidth much lower than that.

[This message has been edited by Wilan Wong (edited 03-07-2001).]

[RobRich]

To better evaluate memory bandwidth, one also must consider the GF-3 supports hidden surface removal, hardware texture compression, hardware z-buffer compression, hardware z-buffer culling, and hardware texture culling. Currently, these options are not optimized (or even enabled in some cases) at the driver level. As better driver support develops, expect these options to supplement memory bandwidth by 15-25%.

Robert Richmond

[Wilan Wong]

I know that NVIDIA is trying to make the memory bandwidth maximum as possible. They should introduce something like Quad Data Rate RAM which would more easily solve the memory bandwidth issue. But that's a long way away, why doesn't NVIDIA move to a 256-bit memory interface or is it just too costly??

[RobRich]

True 256-bit memory is currently way too costly for a retail market video card. A better option would be to move towards a 2x128-bit parrallel memory interface, or better yet, a non-unified memory architecture for each rendering pipeline. After buying 3dfx, one would think nVidia might learn about the positive aspects of non-unified memory addressing.

However, the best option for maximizing memory performance is tile-based rendering. Tile-rendering cards with proper texture compression capabilities offer impressive performance with nothing more than standard 128-bit SD-RAM. Reference the ImagTech/STMicro KYRO PowerVR-3 chipset for example.

Another exceptional tile rendering offering was GigaPixel's GP-1. It offered near GeForce256 performance back in early 1999! The GP-3 was ready for alpha hardware with specs for 25+ sustained GT/s. Most are likely now thinking, "What happened to GigaPixel?" GP was purchased by 3dfx, then 3dfx's intellectual property was purchased by nVidia. Needless to say, GigaPixel is likely a dead project for the near future.

Robert Richmond

[RobRich]

I would enjoy working for nVidia, ATI, or even Matrox within the public/press relations or marketing departments. However, the types of positions I am best suited for tend to require that MS degree (plus several years of experience) you mentioned.

Currently, I'm only 20 years old with an 2-year AS degree in Mathmatics/Computer Sciences. Perhaps one of the days I can obtain one of the above mentioned positions, especially after I ever decide to get back to the studies to complete my BS degree plus a couple of more years tech writing.

Thanks for the positive comments. http://www.sysopt.com/forum/smile.gif

Robert Richmond
Content Developer
www.internet.com (http://www.internet.com)

[Szech]

What I don't quite get, is that if 3dfx owned Gigapixel's intellectual property, and Nvidia now owns 3dfx's intellectual property, then theoretically, doesn't Nvidia have access to Gigapixel's technology? Should Nvidia decide to implement all these technologies, I would expect that they could rule the market.

Then again, from a marketing standpoint, they already do, and don't need to make any new products. Oh well.

[RobRich]

You have the correct idea. nVidia now owns all IP of GigaPixel, which many think would benefit us end-consumers. However, nVidia already has set design programs for the next 24-36 months in place. These development programs have and will continue to cost serious money to fund, thus nVidia will not scrap the projects simply because a better technology comes along. I have heard that nVidia already has a tile-rendering card in development for release in late 2002 or early 2003. Perhaps they will integrate GP technology into these future chipsets.

Same happened with 3dfx. 3dfx purchased GP, but already had development programs for the VSA-100 and next-gen multi-chip Sage/Rampage chipsets. 3dfx planned to offer a GP-2 or GP-3 solution sometime in late 2001, way too late to spur interest in the now closed company.

Robert Richmond

[[H2O]Tman]

**** Robrich you seem to know more than the companies...although im sure they pay their engineers who probably have master's degrees or above to know about the stuff you're talking about. Gee...maybe you should jsut start your own company. You know so much more...