Texel Fill Rate

Fill rate absences?

What we mentioned in our first article on the GeForce 256 was the fact that the texel fillrate was conspicuously absent from anything mentioned by Nvidia. A 480 Million pixel per second pipeline is an impressive number, and Huang mentioned that 480 Mpixels /s is the lowest, not the highest we can expect from the GeForce 256.

So what is the texel fillrate? Well, earlier on IRC, Nvidia's Nick Trianos let slip that the GeForce 256 "has a four-pixel pipeline, and can render 2 textures." Now, this statement would imply that the GeForce could render up to 960Mtexel/sec. This is almost 3 times as many as Voodoo3 3500's 366Mtexel/sec number, and a substantial enough improvement where it might be trumpeted. So what's the deal?

Door number 1

Possibility 1 is that all of the attention has been placed on GeForce 256's geometry calculation engine, and texture fill rate was not significantly improved (at least by the same 2-4x improvement as everything else). Most likely this would mean a 480 Mtexel/sec texture fill rate, nothing worth reporting for this generation of product, especially when S3's Savage2000 will run at 700Mtexel/sec at 175MHz or 800Mtexel/sec at 200MHz.

Discreetly waiting

Another, more likely possibility is that the GeForce 256 does the expected - they process two textures per pixel through four pipes, giving up to 960Mtexel/sec. Now that's something to write home about. So why is it such a huge secret so far? Nvidia might be lying low on the number and waiting for 3dfx to disclose texture fill-rate numbers for their next-generation product, which many rumors have as exceeding 1000 million texels per second (putting it on-par with the announced specs for Bitboys' Glaze3D).

Whatever the reason, we know where the G3Force's strengths lie. Someone asked about texture fill rate at the conference and Jen-Hsun Huang mentioned that fill rate is important but that it's all about the geometry. That's not a surprising reaction, given that geometry acceleration is the big thing that GeForce has over anyone else.

Remember that while geometry acceleration is great for low resolution, traditionally "CPU-bound" applications in today's games, it can't be fully utilized until developers start churning out high-poly count games, and they may or may not be willing to develop for such an obviously superior feature (look at S3TC). Why? Think of the extra work involved, and the need to support 99% of the existing user base, who have zero T&L acceleration.

Fillrate is king…for now

Pixel and texel fillrate will be king for today's relatively low-poly, high resolution gaming. The Q3 benchmarks you see above have very little to do with T&L, and everything with GeForce's superior fillrate. A hypothetical accelerator with 800Mpixel/sec and 1600 Mtexel/sec fillrate might be able to push Q3 1600x1200 up to the 60-80 fps range, and if we can accomplish that, performance bottlenecks will again shift to memory due to the immense amount of data needing to be shuttled back and forth.

Update: 9/2/99, 5:30PM

Nick Triantos from NVIDIA just mailed us with some numbers that clarify some of the above speculation:

In the following snippet of the chat, I wrote we can do 4 pixels, and two textures. I'm sorry if this is misinterpreted, I was typing rather fast to keep up with the pace of messages I was being sent.

GeForce has 4 pixel pipelines. Either they render 4 single-textured pixels per clock, or two dual-textured pixels per clock.

Wanted to clarify so that I'm not the cause of a series of misinformation.

This clarification indicates that the Texel fill on the GeForce256 is indeed 480MTexels/second at a clock speed of 120MHz. That's significantly lower than what Savage 2000 is claiming and far, far lower than what the 3dfx Napalm will likely be capable of.