Nvidia loses patience with TSMC

[Mulciber]

http://www.theinquirer.net/?article=4681

I know some of you guys are real fab buffs.
Considering that nVidia already has a working relationship with UMC for their nv2a, what is the possibility that they could split the load between TSMC and UMC for their nv30 launch, in order to fully ramp up by their intended "fall 02" goals?
How much different could their design rules be that would make it difficult for this to not be feasible?

Thanks

 

[jb]

This could be a huge thing. Any time you have to move a design to another fab, you are basically starting a new design process. You can give them all the process tweaks that the last one had and you will find it could very well require a different set of tweaks in the new fab. For us we when take a design and move to a different fab house its a big deal. We handle it the same as if that design was new so it has to go through all of the qualification process again. Often we have found that a slightly different set of tweaks were required (some times they refer to this a the recipe). UMC process could be a bit better or worse that TMSC so its really hard to say how easy it would be to port over. And given the complexity of the NV30 is not a walk in the park....

 

[Guest]

Yes this is pretty bad news for nvidia. Hopefully(but I doubt it.) they can make the transistion quick enough to stay in the game. nv30 is a long way off if this is the case. Even further then I originally thought.

 

[MfA]

This cant be a spur of the moment thing just for .13u, the shifted volume is way too big for that. Hell, given the volume remaining with TSMC it might even be they are keeping their highest end parts there for the moment being.

Personally I still say NVIDIA also has long term strategic issues in mind for making the shift, the AMD connection

 

[Tahir2]

I dont understand what is spur of the moment and why is this bad news for NVIDIA.
They are still with TSMC but it has been known for almost a year that NVIDIA would not be relying 100% on TSMC due to the fact that TSMC had not delivered .13 when NVIDIA wanted it. It has been known for a long time that they would indeed outsource and UMC was mentioned.

There is no new news just UMC saying that they expect to win orders (small orders at that) which could help them with revenue as UMC has had a bad time recently.

Er... at least that is what I read into it... there doesnt seem to be any 'read between the lines' type thing here at all.

 

[jb]

When I said huge I did not mean anything bad. I just was trying to say its not always easy to take a complex part from one fab to another...

 

[Doomtrooper]

http://biz.yahoo.com/cc/8/20568.html

According to this..Nvidias CEO is stating NV30 is not taped out yet and won't be ready until the holiday season.

 

[MuFu]

http://biz.yahoo.com/cc/8/20568.html

According to this..Nvidias CEO is stating NV30 is not taped out yet and won't be ready until the holiday season.

This is JetLab's post from Rage3D:

Originally posted by JetLab
http://biz.yahoo.com/cc/8/20568.html

He says the NV30 has not finished taping out and it won't be available until the "holiday season"

Nvidia also warned they will miss revenues by $150m and will breakeven this quarter.

Unfortunately I can't access the stream right now so I'll just take JetLab's word for it; NV30 "has not finished taping out".

What I was wondering is whether he means that they haven't completed the initial tapeout or whether he means that there may be further respins before the ASIC qualifies. It seems to me that within the industry there is some dicrepancy in the exact terminology and some people refer to multiple tape-outs (indicative of several silicon respins) whereas others say "tape-out" and they mean the initial upload to the fab for the first mask.

If we take the first meaning, then things might not be as bad as some people think, although "holiday season" sounds like Xmas to me. :-\

MuFu.

 

[asicnewbie]

> what is the possibility that they could split the load between
> TSMC and UMC for their nv30 launch, in order to fully
> ramp up by their intended "fall 02" goals?

From a practical standpoint, it is not desirable to split the production of a single product between two different foundries (UMC and TSMC.) As others have pointed out, that's a large duplication for the backend design process, to achieve essentially the same single product as you already had (except now you can manufacture it at UMC instead of only TSMC.)

Of course may be good business reasons to pursue a second-source.
1) NVidia believes TSMC cannot supply enough capacity for the NV30

IMHO, it makes more sense to move some productline over to UMC. Hypothetically, over the long term, some of NVidia's older products are migrated to UMC, and no single part is fabbed at two places. (During the transition period, maybe some products can be fabbed at both foundries.)

Without any other concrete info, can't really conclude anything. NVidia has a lot of money, and could very well pursue whatever it wants to, like prototyping every product at both foundries.

Given UMC's smaller capital expenditure budget, I find it unlikely that they can beat out TSMC's development schedule for 0.13u, especially for a design of the NV30's complexity.

http://www.eetimes.com/story/OEG20020730S0005 -
"UMC cuts spending"

 

[Vince]

Without any other concrete info, can't really conclude anything. NVidia has a lot of money, and could very well pursue whatever it wants to, like prototyping every product at both foundries.

I totally agree, they have a very effecient and talent group of engineers with the resources they need in their process and test engineering groups. Thay have spent the time and resources to qualify multiple fabs at TSMC and I can't imagine an entire, abrupt, shift to UMC - especially of their high tier, high complexity designs. nVidia get the amazing yeilds they do at TSMC because of the dedication and resources, over time, they pour into the process and test groups. It's [High yeilds] not something you relearn on a new library on a totally new foundery overnight... hell, even individual fabs at TSMC have slighly diffrent process diffrences that require them to be seperatly qualified by the user (ie. Nvidia).

 

[KimB]

If we take the first meaning, then things might not be as bad as some people think, although "holiday season" sounds like Xmas to me. :-\

MuFu.

Technically, the "holiday season" lasts from Thanksgiving to New Years (i.e. 4th Thursday in November to Jan 1st).

This basically means sometime in december.