7. How has Intel responded to the introduction of T&L? Positive, negative...

Obviously I cannot speak for Intel, so you might try asking them directly for their opinion on this. 3dfx feels that for this next generation of products that it is much more important to concentrate on the problem of limited fill rate and to improve image quality by doing real-time spatial anti-aliasing and adding special effects than to start offloading functionality of the CPU. Because the reality is that games will not be able to take advantage of hardware geometry for quite some time in the future, and in the meantime Intel continues to introduce faster and faster processors.

8. DX7 introduces new effects like Vertex Blending. What is it and is it really
useful? Can you give some examples?

Vertex blending is simply being able to take a given vertex and have it be transformed by more than one matrix. This can be quite beneficial in some cases when trying to model joints and other smoothly connected points on a model. To be honest, however, it's not a feature that we've heard developers asking for so I think we're going to have to wait and see if it gets adopted widely.

9. Bump Mapping: Embossing, Dot3, Environment... which do you prefer and why?

Well, quite frankly all of them have their own merits and drawbacks. I am personally not a bump mapping expert so I'm going to pass on this one…All I know is that none of them seem to be resonating very well with the development community.

10. Cube environment mapping... is it really as good as NVIDIA makes it sound or
do you agree with the problems I pointed out in our article?

I think you nailed it with your assessment in your article. It certainly was no coincidence that none of the demos NVIDIA showed with NV10 had dynamic environments. Since the performance penalty is so severe using cube environment mapping, we believe it will be relegated to very specific special effects features. And, we also believe, those specific special effects can just as readily be done with spherical environment mapping. So, all-in-all, I don't believe cube environment mapping to be adopted in any widespread fashion by the development community.

11. Can T&L scale with the CPU? Will T&L on a P2-300 be the same as on a P3-500?

Well, this is a more complicated question than it sounds because often times faster CPUs are coupled with better chipsets, faster system memory configurations, faster AGP busses, etc. However, if you could truly compare an apples-to-apples P2-300 versus a exactly-equipped P3-500 (same memory chipset, same memory type, same AGP bus speed, etc.) then I think the geometry performance for a full hardware-accelerated T&L engine would remain fairly constant. That is not to say, however, that the overall gaming experience is still not better on the 500 MHz P3 than it is on the 300 MHz P2. There are still significantly complex work that needs to be performed on the CPU, such as AI and physics. What will make the biggest difference on hardware-based T&L is memory and AGP bus speed. It's very unclear to us how well hardware-based T&L solutions will perform on AGP 2x and below busses. Because in order for hardware T&L to offer substantial performance advantages, the vertex data has to be pulled directly from AGP system memory without any CPU intervention. The problem with this is that even triangles which are backfacing and would normally be discard (i.e. "culled") by the driver now get transferred down to the 3D accelerator. So, in effect, hardware T&L chips will actually require more AGP and memory bus bandwidth than chips which perform the transform in software. So, we really feel AGP 4x is almost required for good hardware T&L performance. That is something most consumers don't fully understand quite yet.