walterg74 wrote:

That doesn’t sound right at all.. I think you are confusing the data on one card with the data on two cards in SLI.

walterg74 wrote:

Regarding the “works in 6MB V1 but not on 8MB V2”, why is that? Don’t both have 4MB texture memory..? 😲

In theory and technically, you are right and if you use Glide APIs. Yes, you would load texture A in TMU0, texture B in TMU1 and then you can do alpha/color blending with 2 textures in single-pass. Unfortunately, this type of programming is complicated and requires the concept of discrete texture units with absolute addressing. I vaguely recalled that DirectX 6.1 Direct3D could not do that even though it introduced multi-texturing. No real games would pay the labor price to do this even if they support Glide APIs. So in practice, developers would just load the same textures in TMU0 and TMU1 and do single-pass multi-texturing at different addresses. Heavy Gear 1 supports Glide APIs but it still requires 4MB texture storage and does not work with 2x2MB.

walterg74 wrote:

lost77 wrote:

The RAM connected to each TMU contains identical data. A TMU can not access the others RAM so both need all the data for the frame.

So it is correct to say the 8MB Voodoo 2 has 2MB texture memory even though it has 4MB RAM dedicated for that purpose.

That doesn’t sound right at all.. I think you are confusing the data on one card with the data on two cards in SLI.

No, the Voodoo 2 does not have shared memory. That is also why it has a 800x600 limit, 1024x768 with z-buffering can not fit in the 4MB connected to the framebuffer interface. Each of the 3 main chips has a 64-bit interface to its own memory chips. Yes, that was how they made such a fast card without charging a fortune. It was a simple design.

Look at the bottom diagram on page 12 here: http://darwin-3dfx.sourceforge.net/voodoo2.pdf

lost77 wrote:

Yes, that was how they made such a fast card without charging a fortune. It was a simple design.

I think you've got your historical data wrong about 3Dfx. 3Dfx Voodoo was a complicated design and architecture with separate FBI and TMU chips, multi-TMU configuration and mutli-FBI/TMU configuration. Compared to single-chip 2D/3D combos from S3/ATI/Matrox, that in engineering means complicated. They did charge a fortune for what they offered, Voodoo1/2 were never considered cheap or valued proposition back then, regardless if one was discounting the cost of separate 2D companion card. Nevertheless, 3Dfx did enjoy early success with Voodoo1 because they had no competitors in terms of the performance and the 3D feature sets showcased by Voodoo1, hence established the standards for the PC game industry into the dawn of 3D acceleration. Personally, I do not think Voodoo2 was as successful as Voodoo1 for 3Dfx. When Voodoo2 was introduced, competitors had caught up and closed in on the 3D race. And, we all know how the story ended up eventually.

I would consider it more complicated to design a single chip that incorporates everything and have it all work properly on as little a die area as possible than to separate parts in ASICs and patch it together on the PCB.

Even though the PCB is more complicated when you separate components that does not compare to the insane complexity of creating microprocessors.

Voodoo 1/2 were designed as arcade/simulator hardware first and desktop use was more of an afterthought surprisingly.