NV45?

[Maintank]

My guess, and that is really all we can do is guess, is NV45 in Sept\Oct time frame.

Probably using low K for increased clock speeds and possibly faster ram.
I doubt there will be much arch differences between the NV40 and 45, and the main differences will be lower power consumption and higher clocks.

Wouldnt be surprised to see a 550Mhz NV45 with a single molex connector if they use Low K.

 

[karlotta]

just cant see low-k in there future, and if they do id think it will be way bac in line for TSMC is Full, more like end of year, which is january!. Tho i dont have a clue about contracts.

 

[Maintank]

My understanding is all of TMSC lines are or have the capability to do production runs using the low k process.

Isnt the 6800 currently produced by IBM?

 

[Joe DeFuria]

My understanding is all of TMSC lines are or have the capability to do production runs using the low k process.

My understanding is that it's just 0.13 and 0.09 lines.

Not 0.15 or 0.11.

Isnt the 6800 currently produced by IBM?

Correct, but all reports indicate that nVidia has left IBM, and the NV40 is the last chip they fabbed. (At least until nvidia decides to switch back?)

 

[Bjorn]

Correct, but all reports indicate that nVidia has left IBM, and the NV40 is the last chip they fabbed. (At least until nvidia decides to switch back?)

Perhaps it should be "all rumours" rather then " all reports"

Jen Hsuns talk about lower cost's because of using a cheaper process doesn't really agree with leaving IBM for TSMC. Remains to be seen how much truth it's in that statement though.

 

[karlotta]

Correct, but all reports indicate that nVidia has left IBM, and the NV40 is the last chip they fabbed. (At least until nvidia decides to switch back?)

Perhaps it should be "all rumours" rather then " all reports"

Jen Hsuns talk about lower cost's because of using a cheaper process doesn't really agree with leaving IBM for TSMC. Remains to be seen how much truth it's in that statement though.

it would if it was on FSG at .11

 

[Bjorn]

it would if it was on FSG at .11

Aah, damn, forgot about that. Hmm, interesting.

 

[lyme]

Correct, but all reports indicate that nVidia has left IBM, and the NV40 is the last chip they fabbed. (At least until nvidia decides to switch back?)

Perhaps it should be "all rumours" rather then " all reports"

Jen Hsuns talk about lower cost's because of using a cheaper process doesn't really agree with leaving IBM for TSMC. Remains to be seen how much truth it's in that statement though.

Personally I'd consider nvidia moving from ibm back to tsmc just a rumor. Now while I don't know for sure, there are a number of good reasons for staying with ibm over tsmc:

-ibm only charges for the working chips, and not for every chip even if they are defective like tsmc.
-ibm offers re-design and pre-fab services to help increase the yeild of a chip run.
-ibm is agressivly offering a lower price per chip than tsmc.
-nVidia got screwed with the nv3x at tsmc due to bad yeilds.

If anything I'd expect to see a split production of some chips being done at ibm and others at tsmc.

 

[nelg]

-ibm only charges for the working chips, and not for every chip even if they are defective like tsmc.
-ibm offers re-design and pre-fab services to help increase the yeild of a chip run.
-ibm is agressivly offering a lower price per chip than tsmc.
-nVidia got screwed with the nv3x at tsmc due to bad yeilds.

1) Does not help if you cannot meet demand.
2) Is there any cost for this?
3) If the rumored yield problems are true maybe IBM is pushing for nV to leave? (see your first point).
4) seems to have been Nvidia's problem not TSMC.

BTW Welcome.

 

[Dave Baumann]

-ibm only charges for the working chips, and not for every chip even if they are defective like tsmc.

Previously it was said that NVIDIA had this pricing model with TSMC - whether they'll get it again is another question. However, I question this - fundamental economics must still apply for the fab, and the wafer is still a cost to them. Surely the price of each chip per wafer must still be structured to the wafer costs in the first place - take an extreme scenario of one chip yielding from a wafer, is the fab going to take a hit from that (afterall, the issue may have nothing to do with their process, but rather the customer design) or are they going to structure the cost of that chip round the cost of the wafer?

 

[Evildeus]

Anyone, DB?, got information on the previous yield of Nv chips @ IBM?

DB is certainly right, the question is price per chip vs price per wafer /yield. Both can be best cases depending on the price per chip, price per wafer and yield.

Now it seems that IBM have got difficulties to make money on this particular bisuness lately, is IBM risking more on its price list?

 

[lyme]

-ibm only charges for the working chips, and not for every chip even if they are defective like tsmc.

Previously it was said that NVIDIA had this pricing model with TSMC - whether they'll get it again is another question. However, I question this - fundamental economics must still apply for the fab, and the wafer is still a cost to them. Surely the price of each chip per wafer must still be structured to the wafer costs in the first place - take an extreme scenario of one chip yielding from a wafer, is the fab going to take a hit from that (afterall, the issue may have nothing to do with their process, but rather the customer design) or are they going to structure the cost of that chip round the cost of the wafer?

I belive ibm forces their customers to put their designs through their design testing software to ensure a high yeild. If your yeilds are high I'm sure fabs can swallow the cost of defects.
Either way, if the information http://www.eetimes.com/showArticle.jhtml?articleID=18308753 is considered true (and they don't say who their sources are). If I were nvidia and having "But sources said TSMC's 0.13-micron yields actually range anywhere from 40 percent to 70 percent, depending on the product line." yeilds, I would look for a new or additional fabs.

 

[Dave Baumann]

If your yeilds are high I'm sure fabs can swallow the cost of defects.

The problem is that there have been consistent and repeated reports of customer disatisfaction of yields from IBM;s customer division. IBM addmitted to low yeilds here in one of their recent conference calls.

 

[lyme]

-ibm only charges for the working chips, and not for every chip even if they are defective like tsmc.
-ibm offers re-design and pre-fab services to help increase the yeild of a chip run.
-ibm is agressivly offering a lower price per chip than tsmc.
-nVidia got screwed with the nv3x at tsmc due to bad yeilds.

1) Does not help if you cannot meet demand.
2) Is there any cost for this?
3) If the rumored yield problems are true maybe IBM is pushing for nV to leave? (see your first point).
4) seems to have been Nvidia's problem not TSMC.

BTW Welcome.

1) very true, and I hear IBM was having difficulties at the begining of the year.. nothing recently though.
2) Probably, although I would expect it to be rolled into the cost. It's also a selling feature of using the ibm fabs.
3) I don't think IBM would try to push anyone out, but would rather work at fixing their yeild problems. We should remember that TSMC's low-k process is based on using FSG, and while IBM does use FSG they are working on getting SiLK working as their low-k dialectric.
4) As far as I know both ATI and nVidia do not design their chips by hand, they use automated tools. (unlike intel and amd who do). If the yeild for nVidia chips are so low at TSMC it really is nVidia's problem but not their fault. Companies like nVidia if they have a problem they move foundries where they will not have problems that are not their fault (in theory anyway).


It's also important to remember that while both IBM and TSMC have 0.13 lines, the size and design of the transistors are not the same. There was a quote where someone compared the die sizes for the 6800 and the x800, while the 6800 had ?22%? more transistors it was only a die size of 7% larger.
Then again ATI and nVidia count transistors differently, nVidia counts all of them on the die and ATI excludes some (they haven't been specific about it).

 

[Dave Baumann]

4) As far as I know both ATI and nVidia do not design their chips by hand, they use automated tools.

ATI do some. Basic seems to think that NV30 didn't appear to use any hand design; we don't know about NV40.

If the yeild for nVidia chips are so low at TSMC it really is nVidia's problem but not their fault.

Regardless of whether its autorouted or not, the designs from the semicon can be a cause of any yield problem. Moving fab is not the only answer.

It's also important to remember that while both IBM and TSMC have 0.13 lines, the size and design of the transistors are not the same. There was a quote where someone compared the die sizes for the 6800 and the x800, while the 6800 had ?22%? more transistors it was only a die size of 7% larger.

The quoted figure was 9% and that relates to surface area.

 

[lyme]

If the yeild for nVidia chips are so low at TSMC it really is nVidia's problem but not their fault.

Regardless of whether its autorouted or not, the designs from the semicon can be a cause of any yield problem. Moving fab is not the only answer.

That gives more reason to move to IBM at the time. While TSMC is a great company it is hard to belive that they have better design/testing tools than IBM does. It is only recently that TSMC had announced the bundling of their design tools/methodolgies along with their fab services, while IBM had been offering those from the getgo with thier fishkill fab.

Anyway as fabs go, there really doesn't seem to be a clear winner because of the lack of formal information from the fabs. Which doesn't surprise me as they don't want to loose customers.

 

[rainz]

Biggest problem i believe is memory speed .

Its not going to be much faster before xmass . So while i expect it to be faster than the 6800ultra , i don't believe it will be much faster. Perhaps a smaller jump than the nv30-nv35 was

OCZ will release 650mhz memory sticks .. but dunno if we can apply the same speed for the video cards

RainZ

 

[Joe DeFuria]

OCZ will release 650mhz memory sticks .. but dunno if we can apply the same speed for the video cards

No, that's 325 Mhz DDR, 650 "Effective." Video cards are using 600 Mhz / 1200 effective.

 

[Dave Baumann]

That gives more reason to move to IBM at the time. While TSMC is a great company it is hard to belive that they have better design/testing tools than IBM does. It is only recently that TSMC had announced the bundling of their design tools/methodolgies along with their fab services, while IBM had been offering those from the getgo with thier fishkill fab.

Ho do you reconcile this with the fact that only one of the high end graphics ASIC manufacturers have been bemoaning the yield rate at TSMC whilst the other has consistently praised it so far?

 

[lyme]

Ho do you reconcile this with the fact that only one of the high end graphics ASIC manufacturers have been bemoaning the yield rate at TSMC whilst the other has consistently praised it so far?

Did ATI praise the yeild rate? or just haven't said anything about it?

Of course the people at nVidia did have to blamestorm with the nv3x products being so craptacular. Although I don't see why they would blame them or switch fabs if everything was just 'fine'.