Best solution to bandwidth problems in video cards?

[Muerto]

Hello people,

What do you guys think the best solution to the memory bandwidth problems in video cards like the GF2 is?

Personally I think that the memory bus width has to be increased to 256 bits at least. That would give a GF2 over 10 GB/s of bandwidth. We could then get a lot closer to the theoretical fill rate of 1.6 Gtexels/s.

I hope that nVidia doesn't just put 200 MHz DDR SDRAM on the NV20, that's only 6.4 GB/s, as much as the Radeon256. The NV20 wouldn't really be worth buying if that were the case.

I think that Bitboys had the right idea of using embedded DRAM in their Glaze3D. It had a 512 bit memory bus! Unfortunately we wont be seing that anytime soon.

Anyway, thanks for your input.

 

[blade47]

I think I read somewhere that Nvidia's next chip will use 256bit.

 

[IvoryGrail]

But also keep in mind that the NV20 might use a hidden surface removal technique which would greatly reduce bandwidth. Add the rumors of Z-buffer compression and others, 6.4 GB/sec is not all bad.

 

[Goi]

Memory isn't going much faster overall, so new algorithms and techniques have to be applied to reduce memory bandwidth consumption.
A 256-bit memory bus isn't really feasible for external memory, as it could be really expensive. Embedded DRAM could alleviate some of the problems we have, but its potentially only a temporary solution as you can't embed too much memory into the core without increasing the die size and cost.