S3 Interview

Fill rates and T&L

I first asked Paul and Raja about the difference between pixel fill rate and texel fill rate. The GeForce does 480Mpixels/s, but only 480Mtexels/s while the Savage2000 can do about 700Mtexels/s, but only 350Mpixels/s. Which fill rate is more important in today's games?

Raja jumped all over this one. He explained that the texel number is more important now because there are a lot of multi-texture games out there such as Shogo, Quake 2, and Quake 3. In general, all first person shooter type games will use multi-texturing, and the Savage2000's 700Mtexels/s fill rate will be much more important than the 350Mpixels/s.

Satisfied with Raja's answer I moved on to the Savage2000's T&L implementation. I first asked about how well the Savage2000 scales on different processor speeds. Would there still be a framerate difference between an Athlon 500 and an Athlon 800(KryoTech) in Quake 3 even though the T&L has been moved on chip?

Raja answered in the affirmative. Yes, the processor will still scale up with faster processors. In Quake 3, the processor definitely has other duties to work on besides pushing out the triangles. In fact, the processor only spends about 40% of the time in OpenGL.

How many triangles?

Then I asked how many triangles could the Savage2000 sustain compared with the GeForce's 15 million triangles/s figure. Raja didn't have any sustained numbers, but he offered a safe answer: between 10 and 20 million triangles per second.

I then asked about the purpose behind real-time load balancing between the CPU and the T&L unit. Calbear had asked Nvidia why the GeForce didn't support real-time load balancing, and they said that it wasn't needed or even recommended because sharing the triangle load with the CPU would slow everything down.

Raja explained that the CPU would be performing other tasks while the Savage2000's T&L churned through the triangles. It sounds like we just had a misunderstanding over the meaning of real-time load balancing. S3 never meant to imply that the CPU could split triangle time with the T&L unit. S3 marketing just added it to the Savage2000 presentation as a special feature when they didn't even need to mention it.