NVIDIA Kepler speculation thread

[UniversalTruth]

You're forgetting Nvidia has no choice but to price lower than their performance dictates because AMD has the stronger brand

Yes, sarcasmwise. I realise that I don't remember something that doesn't exist.

The sad thing for all customers is that it will indeed take quite a long time for NV to calm down these whooping supply-demand issues. And then- to enjoy sane prices. They will simply either EOL the product by then, or simply reduce (with respect to the potential maximum) the manufactured units amount just to *protect* the price.
See how that's evil from any other POV but theirs.

 

[Alexko]

That's hardly an achievement. What do they think? That if their product is priced at 300 and 500 demand would be one and the same?

NV can do it too. Put the 2000 price tag for the 680 and almost all of them will be widely available even with no need to manufacture more.

But they are not doing this. They keep relatively low pricing, sell everything and prevent AMD from doing additional sales.

What are you talking about? The 680 is $20~50 more expensive than the 7970, and the 690 costs $1000.

 

[UniversalTruth]

What are you talking about? The 680 is $20~50 more expensive than the 7970, and the 690 costs $1000.

And so what? That is a value which shows how much NV brand is stronger and why AMD desperately needs lower pricing to justify anyone's buy.

 

[trinibwoy]

Charlie may or may not be right about shipment figures, but one of his points is hard to argue with: only NVIDIA is openly complaining about 28nm yields, and Thomas Seifert did say that AMD was able to meet demand for their 28nm products last quarter.

nVidia may simply need more capacity from TSMC than AMD does or are just more vocal with their criticisms. It's been known for some time too that they're aggressively pushing a mobile strategy with GK107.

I don't see any evidence of particularly high volumes of AMD 28nm parts. Like it or not but the Steam survey is currently the best proxy by fr for actual sales figures of 28nm chips. AMDs parts aren't showing up in any great quantities and they have lots of 28nm SKUs out based on 3 different chips starting at very affordable prices ($100).

 

[Dave Baumann]

The facts are that TSMC does not have 28nm wafer production yields anywhere near the amount required to fulfill the orders currently being placed by AMD, nVidia, Qualcomm, and many other companies. FAB 15 is still a ways off from volume production and the only FAB currently producing 28nm wafers will not be at full capacity until early 2013.

Good suppliers do not capacity plan in a vacuum, the do it in conjunction with the customers. Planning the capacity requirements is in the customer hands though and if they poorly plan before hand (by not correctly factoring the yield expectations in the planning phase) or see lower then expect yeilds as the product comes out then they will end up in short supply.

Not everyone is unhappy with the 28nm supply situation.

 

[Dave Baumann]

That's hardly an achievement. What do they think? That if their product is priced at 300 and 500 demand would be one and the same?

All three products had launched in calendar/financial Q1, so we had a stack starting at $109.

 

[Andrew]

Good suppliers do not capacity plan in a vacuum, the do it in conjunction with the customers. Planning the capacity requirements is in the customer hands though and if they poorly plan before hand (by not correctly factoring the yield expectations in the planning phase) or see lower then expect yeilds as the product comes out then they will end up in short supply.

All the news I can find from reputable sources seems to be pointing toward TSMC not having enough capacity for their clients. Its becoming such a problem for Qualcomm and nVidia that they both just requested supply priority from TSMC and were approved.

http://www.digitimes.com/news/a20120509PD211.html?mod=2

Not everyone is unhappy with the 28nm supply situation.

Qualcomm and nVidia, TSMC's two largest customers for 28nm are both pisssed. And considering how simple the new snapdragon chip is yet qualcomm still is receiving low quantities.... it all but confirms what i've said.

http://www.electronicsweekly.com/bl...or-blog/2012/04/dont-upset-tsmc-was-mike.html

 

[3dcgi]

When a new process is ramping you need to have accurate sales/yield targets very early and acquire the capacity from TSMC. If you underestimate and TSMC is sold out they won't be able to accommodate you and increase production in a timely manner. Maybe this has happened to Qualcomm and Nvidia.

The 40nm rollout coincided with a recession and companies didn't secure enough capacity due to fear so TSMC didn't have availability when demand was stronger than expected. The cycle repeats...

 

[UniversalTruth]

That statement makes any point you are trying to make "null and void".

Don't blame Charlie, he's a good guy . And the site is not accidently called SEMIaccurate.
Someone has a very nice opinion about him.

Charlie tends to pull things out of his arse usually, and if you've noticed his so called articles have tags like "humor", "opinion" and "rumor".

http://www.techpowerup.com/165707/TSMC-Gives-NVIDIA-Priority-for-28-nm-Manufacturing.html

Sorry.

 

[Blazkowicz]

Intel is happy about TSMC production woes, now they can sell Atom on cell phones. (but can I run XP or MS-DOS on it, that question was left unanswered)

 

[AlphaWolf]

Source for that techpowerup article is digitimes. They have a track record akin to charlies.

 

[Andrew]

When a new process is ramping you need to have accurate sales/yield targets very early and acquire the capacity from TSMC. If you underestimate and TSMC is sold out they won't be able to accommodate you and increase production in a timely manner. Maybe this has happened to Qualcomm and Nvidia.

I think you completely missed the point here.

The problem is with TSMC, who is not able to meet the demand for 28nm wafers.

Graphics card shortages to see improvements in late May
May 3, 01:20
AMD and Nvidia, impacted by Taiwan Semiconductor Manufacturing Company's (TSMC's) fully-booked capacity, had rather weak shipment performance in the first quarter; however, as more capacity will be gradually released by TSMC, shortages of 28nm graphics cards are expected to improve in late May, according...

The 40nm rollout coincided with a recession and companies didn't secure enough capacity due to fear so TSMC didn't have availability when demand was stronger than expected. The cycle repeats...

Actually it was TSMC who had all sorts of problems with the 40nm manufacturing process, resulting in very poor yields the entire second half of 2009. This is why ATi HD5xxx gpu's were extremely hard to find and nVidia's Fermi had to be delayed because of its complexity and size... and horrible yields.

 

[Dave Baumann]

I think you completely missed the point here.

The problem is with TSMC, who is not able to meet the demand for 28nm wafers.

He's not missing the point. Demand is planned. If, all of a sudden, when the products are shipping someone is saying that demand is not being met then it is more likely that something has gone awry with the planning (i.e. it was either not planned accurately or there are yeild issues).

Once again, not everyone is complaining.

 

[trinibwoy]

Dave why are you mixing yield issues with underestimated demand? They aren't related at all. If there are yield issues then it can very much be a TSMC problem.

Do you have any info to indicate that TSMC's clients underestimated demand for their products? It would be rather stupid to blame TSMC for not providing more capacity than was requested so I have to assume that's not what's happening here.

 

[Alexko]

Dave why are you mixing yield issues with underestimated demand? They aren't related at all. If there are yield issues then it can very much be a TSMC problem.

Do you have any info to indicate that TSMC's clients underestimated demand for their products? It would be rather stupid to blame TSMC for not providing more capacity than was requested so I have to assume that's not what's happening here.

AMD isn't complaining about yields either. If AMD's yields are acceptable and NVIDIA's aren't, how likely is it to be TSMC's fault?

 

[Dave Baumann]

Not sure what is not clear here. A supplier plans capacity with the input from its customers. If the customer planned their wafer requirements poorly or has lower than expected yeild they are going to need to raise their demand beyond the original planned requirements and the supplier is not going to have capcacity (and hence the customer complains of supply).

With regards to a foundry affecting yields, other than specific excursions (which are one off events), then the main thing that affects yeilds is the defectivity profile of a process. There is a is a roadmap for defectivity, which should be accounted for in the original planning of the wafer needs by the customer.

 

[Gubbi]

Dave why are you mixing yield issues with underestimated demand? They aren't related at all. If there are yield issues then it can very much be a TSMC problem.

If a wafer has X dies, and you expect demand for Y chips, and you 'know' from experience yield will be 45% and not 90% you order 2.2 Y/X wafer starts instead of 1.1 Y/X.

Cheers

 

[trinibwoy]

AMD isn't complaining about yields either. If AMD's yields are acceptable and NVIDIA's aren't, how likely is it to be TSMC's fault?

AMD not complaining is an indicator of pretty much nothing....

 

[trinibwoy]

If a wafer has X dies, and you expect demand for Y chips, and you 'know' from experience yield will be 45% and not 90% you order 2.2 Y/X wafer starts instead of 1.1 Y/X.

Cheers

You cannot "know" anything about a new process based on past experience with other processes. You can only hope for the best.

Not sure what is not clear here. A supplier plans capacity with the input from its customers. If the customer planned their wafer requirements poorly or has lower than expected yeild they are going to need to raise their demand beyond the original planned requirements and the supplier is not going to have capcacity (and hence the customer complains of supply).

I hear what you're saying but you're still entangling poor planning with bad yields. Part of the planning process is to make assumptions of what your supplier can provide. If the supplier significantly underperforms you're screwed no matter how well you planned.

 

[Dave Baumann]

You cannot "know" anything about a new process based on past experience with other processes. You can only hope for the best.

No, you can do a lot more than that. If that was NVIDIA's attitude then I can see why they would be complaining now.

Part of the planning process is to make assumptions of what your supplier can provide. If the supplier significantly underperforms you're screwed no matter how well you planned.

How, exactly, do you think the supplier can underperform?