GK110 was shipped to Cray in June 2012. It was ready, TSMC wasnt.
Thats exactly what I am saying. Yields matter especially on very large dies. Production volume matters. Apple going with 100% TSMC for A10 / A10X means that others are going to be lower in priority for wafer allocation and going to fight over whats left.
In the case with nVIDIA, yields don't matter when the margins on these GPUs will be x5 to x10 to what you'll get in the consumer space (or more). It could be terrible for all we know but they have orders to fill and deadlines to meet for various supercomputing projects this year. That is why they are pushing the big one out. I suspect that itd be lucky for us consumers to see a Titan version of the big Pascal this year due to yields and priorities.
But definitely the thing has been taped out, and I wouldn't be too surprised to see this launched as a Telsa product within this year.
Seems like they've already given Cray, a major supercomputer vendor, a "commitment."
Kinda ruins the hypothesis that GP100 is coming in 2017, doesn't it?
Thats exactly what I am saying. Yields matter especially on very large dies. Production volume matters. Apple going with 100% TSMC for A10 / A10X means that others are going to be lower in priority for wafer allocation and going to fight over whats left.
Let's say a 16FF+ wafer is $8500. How many good dies per wafer do you need to achieve 55% gross profit margin if you're selling GPUs for $1K/each?
What % yield does that work out to?
Assume die size of 450mm^2.
Seems like they've already given Cray, a major supercomputer vendor, a "commitment."
Kinda ruins the hypothesis that GP100 is coming in 2017, doesn't it?
What do you mean by $ 1K each? The bare die?Let's say a 16FF+ wafer is $8500. How many good dies per wafer do you need to achieve 55% gross profit margin if you're selling GPUs for $1K/each?
What % yield does that work out to?
Assume die size of 450mm^2.
Is that the rumor or just a straight estimation of transistors for a traditional max GPU size of 600mm^2.Rumored 18B transistors for big Pascal. The current 28nm GM200 is ~8B.
The die size of big Pascal would be ~600mm2 given published densities at TSMC 16ff.
Let's say a 16FF+ wafer is $8500. How many good dies per wafer do you need to achieve 55% gross profit margin if you're selling GPUs for $1K/each?
What % yield does that work out to?
Assume die size of 450mm^2.
Let's say a 16FF+ wafer is $8500. How many good dies per wafer do you need to achieve 55% gross profit margin if you're selling GPUs for $1K/each?
What % yield does that work out to?
Assume die size of 450mm^2.
First Kepler Tesla cards were based on GK104, let's not jump to conclusions
EDIT: Good find though, definitely should be a new Titan card by the end of the year!
We're not going to get Titan this year. Once again we'll get mid-range selling as the high-end from both camps. Foundry reality is setting in.
Wouldn't that just lead to no retail units soon, maybe for the year?TSMCs Q4 2015
16nm FF+ was close to 4% of its revenue with 20nm contributing close to 20% and 28nm contributing another 28% of Q4 2015 total revenue.
Since Apple will use the same process that NVIDIA will use and since Yields are not as high as 28nm, the available 16nm FF+ wafer volume is low. NVIDIA will have hard time allocating all the wafers they would like in 2H 2016 for Pascal, for both retail and professional dies.
When you have low wafer volume AND lower yields, creating large dies will only make things worse. And if you have a target volume to meet for the CRAY servers, things will not be easy for those big dies to be available for the retail market.
The problem is that you can't just take the yields of smaller chips and apply it directly to larger chips. The larger the die you try to harvest from any given wafer, the greater the chance that any particular die will have a defect in it thus lowering yields. Here's a graphical representation of the situation:Yeah, no way 16FF was only 4% of their revenue in Q415. If they actually claimed that in financial reports, the 20% they are claiming for 20 nm is mostly all Apple, both the A8/A8X/A9X and the portion of the A9 they got.
Yields really should be Not Terrible. It'll still be super expensive though.
BTW, apparently for a 400 mm2 die, you would get 147 potential good die on a typical 300 mm wafer. So at 70% yield (for both good and cut) you would get 100 maybe? You'd be at $85 already just for the chip itself.
Wouldn't that just lead to no retail units soon, maybe for the year?