NVIDIA's Fermi GF100 Facts & Opinions

[S0ULphIRE]

1.5Gb of ddr5...mmmm

 

[Wildside]

the biggest grip of the Fermi would be heat. I'm sure it will do better than the 5870 n match up with the 5890.

 

[Fat.Clown]

for all i know, it could be cool as winterfresh

 

[Apokalipse]

the hd45870 and hd5970 are having yield issues as well it isn't nvidias fault it is tmsnc 40nm prodution fabb. from my understanding the yield issues are improving. its sad just look at xfx they barely sell nvidia cards anymore i cant even buy a ndvidia gtx295 on ncix but who cares i want an hd5870 anyway lol

Cypress is getting at least 60% yields ATM (probably closer to 70%)

Two things that help Cypress yields are:
* it has about 2/3 the transistor count, and is a smaller die, so less chance of manufacturing defects per die

Mind you, it isn't entirely bad to use a large die on a new manufacturing process. Because there are going to be more manufacturing problems with a bigger die, you can learn more about what issues it has specifically.

AMD tested TSMC's 40nm manufacturing process using RV740 (4770 GPU), which has a 137mm² die size, compared to Nvidia's G210 and G220 GPU's (57mm² and 100mm²) on TSMC's 40nm manufacturing process.

As a result, RV740 didn't have particularly good yields, and it was hard to find 4770's for sale. But from RV740, AMD was able to learn more about TSMC's 40nm process and what bugs it had.

Specifically, AMD discovered that the Vias (interconnects between curcuit layers) weren't up to scratch. So to resolve the issue, they decided to double the number of Vias for redundancy. There is a slight increase in die size by doing so, but AMD can afford to do so with Cypress.

Not only that, but the variation in transistor dimensions is particularly bad, causing high leakage with some transistors and poor performance with others.
I'm not sure exactly how AMD's engineering team modified RV740 to fix most of the issues with channel length, but they did. And they carried what they learned over to Cypress.
AnandTech: The RV870 Story: AMD Showing up to the Fight

Now with GF100, Nvidia has had to try and wrestle these same issues. And with less experience on TSMC's 40nm process, and a much bigger die, the problem is only exacerbated. And Nvidia needs to try and resolve them in order to get power consumption and performance to acceptable levels.

In order to have better performance against Cypress, it will be difficult to get power consumption down and yields up so that it won't be too expensive.

Of course Nvidia is likely to do whatever it can to take the performance crown, so how much power it uses and how expensive it will be depends on Nvidia engineers to design GF100 to work better with TSMC's 40nm process.
AMD has done really well to design Cypress for low power consumption and high performance per watt, so it looks like that's one area where Cypress will be ahead.