Nvidia still reporting 40nm probs to analysts

[Gerard]

Wonder why ati can't make use of one of amd's fab plants, or would more tooling be required for gpu's?

 

[mmj_uk]

I doubt ATI would be paying for lost chip due to yields, I would have thought ATI are turnkey with TSMC, ie they hand over the design a say sell me x number, TSMC are probably the ones taking the hit.

As D.P. explained it is AMD/Nvidia that will take any losses, for that reason they usually sell failed/imperfect dies as lesser models (5850 for example) to reduce their losses, dies that don't even meet those grades are no good to them at all.

Fermi supposedly has a billion more transistors than 5870 so NVidia will want to ensure the process is mature before going ahead with mass production, as bad as it looks they are just sitting back while AMD suffer the early 40nm teething troubles.

 

[Gills]

Is TSMC the only place the green/red side can use ? Is there no other manufacturing plants available to them ?

 

[mmj_uk]

Is TSMC the only place the green/red side can use ? Is there no other manufacturing plants available to them ?

From what I can tell TSMC are the leading foundry, if they can't make 40nm parts then there's probably not much hope for the others yet.

 

[Gills]

From what I can tell TSMC are the leading foundry, if they can't make 40nm parts then there's probably not much hope for the others yet.

Ok thanks for that.

 

[drunkenmaster]

It's more like, NVIDIA is lucky TSMC can't supply enough to AMD/ATi, or am I seeing this wrong?

Yup

I wonder if the yield issues will get miraculously fixed once fermi arrives ...... one for the conspiracy theorists.

nope, it won't, and again I'll point out yields isn't refering to one thing. Theres wafer yield, the physical number of cores that work from each wafer processed and theirs capacity/plant yield, how many wafers it can actually push from one side of the production line to the other. One has zero inpact on the other, you can have a wafer with 1% yield with only 1 working core per wafer, but the plant could be pushing 750million wafers through a day. Or you could have 80% wafer yields, 80 cores per wafer, and the plant is only able to push through 100 wafers a day.

The best situation is huge yield on both area's, the reality is no plant can make infinate numbers of wafers every day, because it takes time. Nvidia and AMD are complaining as TSMC can only make 9k wafers a month, their chamber missmatch problem meant they were down to 6k wafers a month for 2 months. Even back to 9k where they are, its not huge, they really need to move more plants over to 40nm but they haven't yet. Both companies together putting out lots of 40nm really want TSMC to be making 20k wafers or more a month.

I doubt ATI would be paying for lost chip due to yields, I would have thought ATI are turnkey with TSMC, ie they hand over the design a say sell me x number, TSMC are probably the ones taking the hit.

Yes they are, they pay a flat fee per wafer, it doesn't remotely matter if 1 core or 50k cores come off that wafer, though obviously TSMC could adjust their pricing if they new it was that profitable. AS it stands TSMC charge around $3500 for a 55nm wafer, and $5000 for a 40nm wafer.

Its VERY likely that the 6 weeks of wafers lost when TSMC found out about the chamber missmatch(it takes 6 weeks start to finish to produce a wafer, they only found out about it when the first cards came off that particular line, however every wafer through it over that 6 weeks was lost also) they would eat the cost of that themselves, we know thats around 4.5k wafers, which is over $20mil of silicon. They've probably given AMD some deal to help make up for it, discount, promise to put through some extra production at their cost.


The yields at 40nm suck compared to 55nm, Fermi's yield is apparently disgustingly bad, the 58XX aren't supposed to be great at all and as the cores get smaller yields go up, so the 5770 has better yields than something twice its size.

Plant yields are back up, but need to be higher(realistically by adding more fab, you can't turn a 9k wafer fab into a 50k one,), wafer yield is horrible with terrible leakage issues, it seems unlikely to improve dramatically, ever. Its a fairly old process by now, the biggest yield improvements normally happen quickly, after this amount of time it would take a fundamental change in their 40nm process to improve it, like adding SOI or HKMG, either would require both companies to redesign their cores, tape them out, test them then produce them(a 3month cycle at best). Best bet for improved yields are a non crap 28nm process next year. We won't see £100 5850's till 28nm, this generation will NEVER achieve price parity with last gen even with massive competition the cores simply cost more to make due to yields.

 

[drunkenmaster]

From what I can tell TSMC are the leading foundry, if they can't make 40nm parts then there's probably not much hope for the others yet.

yup, right now theres TSMC, Semi Chartered(a singapore owned group of fabs brought out by Global several months ago) who aren't up to speed with the high end stuff till Global invest and re-equip the fabs, best case scenario is next year and 28nm production.

Those are existing plants, then Global are half way to building basically the single best fab in the world, not ready till 2012, though they have AMD's dresden fabs which are very good fabs. Supposedly AMD might be producing next gen parts through Global rather than TSMC, I've no clue where, at the Semi Chartered fabs, at the AMD dresden fabs with spare capacity, at a couple other places.

TSMC should improve though, like any other area they get better with competition, TSMC do the bare minimum to keep going because, no matter how much they screw up, (65, 55, 40nm over what 4 years) theres no where else for AMD/Nvidia to go for GPU production, both gpu makers simply had to take what they were given. With Globals introduction and massive investing, buying LOTS of fabs and having a lot of links with R&D guys with a huge group working together(IBM, AMD and several others).
Well the best thing is TSMC have been forced to quite dramatically increase their R&D spending, they have to screw up less or they'll be out of business, so they'll invest more, spend more and have less screw ups which is good for all of us no matter where the cards end up actually being built.

Samsung also announced they were going to massively expand into the foundry business, they've grown from something like a 200million to a 1billion foundry business in the last couple years alone, they've stated their intention to long term want to compete with TSMC. So expect Samsung to invest heavily aswell, though its unclear how much they'll spread out chip wise, small tiny chips like Arm/Tegra stuff, or fully fledged company with different fabs for anything, including full on GPU designs. Again the increase competition is great for us all. But till Global/Samsung kick into gear we've got a useless 40nm for this generation, next generation is too close to call if Global can be ready in time, the generation after that all 3 big players could be competiting for contracts on the latest and greatest manufacturing process's to rival even Intel's production tech. Interesting times ahead.

 

[tbh]

Thanks for that excellent post drunkenmaster, I was concerned after I heard that TSMC had cancelled their 32nm process however it appears that there is reason to be optimistic regarding competition in the fab business. I guess nVidia must agree since they have stated that they will be sticking with TSMC rather than moving to Global?

 

[LoadsaMoney]

how the hell amd missed the window of opportunity if they released the cards first , it's other way around tbh

Yeah very odd, as the latest score is :- ATi 2 million - Nvidia 0