Never let it be said that facts will get in the way of a fun story - Mike Magee
We can clearly state that this chip is going to be built using 0.13µ (micron) marchitecture and IBM already said that it will play with this chip to make it work for Nvidia, so no surprises there. Nvidia is unlikely to dabble with .11µ architecture yet, we'd say.
This marchitecture and its cost is limited to a maximum 200 millions of transistors in order to earn any money from this wafer. We asked around and learned that the new chip likely won't be eight pipelines x two TMUs, as this isn't enough to keep Nzilla in the game.
Sixteen pipelines capable of handling FPU information and one TMU is impossible at this time we are told, whioch leads us to speculate the next Nvidia part will likely use 12 pipelines.
Twelve pipelines is possible despite it interfering with the nice progression of 2,4, 8 and 16.
The NV40 will be a native PCI Express chip with some kind of bridge to make it work in current AGP boards that will co-exist for quite some time.
Performance wise, it's easy to assume that it will be faster than FX 5900 Ultra as it has to be. Historically, each generation of Nzilla chips is around 30 per cent faster - no more, and with two versions which the boys prefer to call Ultra and Non Ultra rather than Faster and Slower.
As for memory, it's impractical to use more than a 256 bit interface since it would cost you a fortune and these £399 cards are already very expensive. We're not sure which direction Nvidia will take with the memory.
We expect it will use DDR II, since it seems to be the right time and they way to go as GDDR 2 is still a little too early in the hatching.
We understand the NV40 hasn't yet taped out, but in order to compete with ATI's Loki chip, we'd expect a Computex or a November introduction.
As for the name, we are almost 100 per cent sure it will be Geforce something since Graphzilla simply doesn't have any other choice. We would frankly be surprised if it was changed. µ
L'INQ
Geometric progressions