Yesterday I said I was chasing more info on the NVIDIA upcoming G80 chip. The more I dug, the more my head hurt, nothing concrete, so take this with a pretty big grain of salt. In any case, there is some interesting info here. The first thing is I was a little off on the dates, the technology day now seems more likely in late October with a launch in early November. What's a few weeks between friends? If they have not gotten back the latest silicon though, it is going to be really tight to make that schedule, chips take time to fabricate. What they will talk about is the odd part. First is the arrangement of the chip, physically we are hearing that it is a huge 400mm die. Ouch... - The Inquirer

NVIDIA's upcoming G80 mystery starts thickening

One of the reasons it is so big is the whole dual core rumor that has been floating around. The G80 is not going to be a converged shader unit like the ATI R500 or R600 chips, it will do things the old way. Some people are saying that it will have 96 pipes, split as 48 DirectX 9.0 pipes and 48 DirectX 10. While this may sound like a huge number, we are told if this is the case, it is unlikely that you can use both at the same time. Think a 48 or 48 architecture. That said, I kind of doubt this rumor, it makes little sense. In any case, NVIDIA is heavily downplaying the DirectX 10 performance, instead shouting about DirectX 9.0 to anyone listening.

If the G80 is more or less two GeForce 7900's like we hear, it should do pretty well at DirectX 9.0, but how it stacks up to the R600 in DirectX 9.0 is not known. The R600 should annihilate it in DirectX 10 benches though. We hear the G80 has about one third of it's die dedicated to DirectX 10 functionality. One of the other interesting points that surfaced is much more plausible but still odd, and that is memory bandwidth. As far as I know, GDDR3 is only available in 32-bit wide chips, same with the upcoming GDDR4 memory. Early G80 boards are said to have 12 memory chips on them, a number that is not computer friendly.

Doing a little math, 12 times 32-bit gives you a 384-bit wide bus. On the odder side, if you take any of the common card memory capacities, 256MB and 512MB and divide them by 12, you end up with about 21MB or 42MB chips, a size not commonly found in commodity DRAMs. If 12 chips are destined for production, you will end up with a marketing nightmare in DRAM capacities, 384MB or 768MB amount of memory is not a consumer friendly number. While it could just be something supported by the GPU and not destined for production, you have to ask why they may need this. At least to use it sounds a bit strange.

The thing that makes the most sense to me is that with all the added pipes, they needed more memory bandwidth. To get that, you can either go faster or go wider. If you look at the cost of high-end GDDR parts, faster is ugly. If you look at the yields of high-end GDDR parts, faster is ugly. Drawing more traces on a PCB to go wider is far less ugly. If NVIDIA needed the bandwidth that you can only get from a 384-bit wide bus, then the odd memory size may simply be a side effect of that. In any case, the G80 is shaping up to be a patchy part. In some things, it may absolutely scream, but fall flat in others.

The architecture is looking to be quite different from anything else out there, but again, that may not be a compliment. In either case, before you spend $1000 on two of these beasts, it may be prudent to wait for the R600 numbers.