14nm GF volume production with low double-digit yields? Apparently yes.

[blastingcap]

People keep saying this, but has there been any official word from AMD to say no Excavator(or anything else) for AM3+? Or are people just going on rumours?

Common sense and reading between Su's lines.

 

[Arachnotronic]

Globafoundries denying 14nm delay rumors:

http://blog.timesunion.com/business...of-delay-in-samsung-partnership-untrue/62956/

Our Fab 8 ramp is on track and we have yielding customer product on our 14nm technology.

Gotta love it. "On track" and "yielding." Completely worthless words without additional context.

 

[Rezist]

Before moving to Taiwan I lived in Pennsylvania (one of the 50 states in the USA for readers who might not know), and will live there again someday. In the meantime I have family that continues to live in Pennsylvania, specifically on farmland, multiple in fact.

And they are smack in the middle of the fracking fruckus, well aware of the horror stories (which are not stories, they truly are a reality) of fellow farmers who can no longer feed their dairy cows the grass that grows in their fields or allow their animals (chickens, pigs, etc) to drink the water that is pumped from 100+ year old wells on their farmland because of what happened literally the month the frackers moved in and "leased" the surrounding areas for fracking.

When it comes to fracking the risks and damage is clear to me from a first-person personal level, not from a distance by reading articles about it on the internet or from watching news reports on TV. Billions of people can easily be swayed to be apathetic to the issue, happy to save a buck or two on gas versus worrying about preserving the very land they never see with their own eyes but which creates the food they need to eat tomorrow.

This is nowhere more self-evident than the farmers who do accept the fracking leases themselves. They do it for greed because they too would rather have a dollar in their pocket today and not worry about tomorrow until tomorrow gets here.

IMO it is a loosing battle. Humanity is simply too stupid and too selfish of a species to ever put the benefit of the many above the benefit of themselves. We will extinct ourselves, no question, merely a question of what else we will take down with us in the process.



The families that control mubadala and the oil wealth are not worried about themselves and their present day wealth. They are worried about what their families will be doing in 100+ years when they have nothing left but their accumulated warchest and nothing left to create jobs in their country outside of hoping for a sand exporting business.

In many ways they are a lot more forward looking than many western nations who are doing little more than creating a large debt for their future generations with little concern for seeding industries that will create stable long-lasting job growth.

We all can't become Uber drivers. Somebody has to be willing and able to create the products that the rest of plan to kick back and consume while living off our pensions or retirement savings

This is interesting to read, I work in the oilfield in northern Canada where fracking won't work due to viscosity and were way too deep for oil sands (open pit or SAGD) and we go pretty hard lately with SAGD horizontals and just large multi-horizontal drills and new technology like polymer injection. These all have there own environmental issues but contamination is not one of them (lost fresh water underground primarily and CO2). We were set to go ahead with Nuclear reactors to clean up produced water and make steam from it but were shutdown by lobby groups (environmental). We had to get Keystone XL built faster then the fracking boom but it got blocked and will probably never go ahead now. It was pretty clear the US gov just wanted to get fracking going hard and did not want to purchase any more Canadian made oil. Now we sit with excess production and no reasonable way to get it to consumers.

 

[el etro]

Gotta love it. "On track" and "yielding." Completely worthless words without additional context.

And the 14nm delay rumor worth what? Worth nothing. Digitimes don't have a very good track with their rumors.

 

[cbn]

Just out of curiosity, what would AMD's expected yields be for 32nm and 28nm at GF today?

32nm for Vishera 315mm2 die size
28nm for Kaveri 245mm2 die size
28nm for Beema/Mullins ~101mm2 die size

Percentage wise does anyone want to wager a guess for each of those different chips I've listed above?

 

[Abwx]

Just out of curiosity, what would AMD's expected yields be for 32nm and 28nm at GF today?

32nm for Vishera 315mm2 die size
28nm for Kaveri 245mm2 die size
28nm for Beema/Mullins ~101mm2 die size

Percentage wise does anyone want to wager a guess?

Vishera 8C should have good yields given the process maturity, about as good as the 245mm2 Richland APUs as the the former s density is much lower, the effective sensitive area is the same for the two chips.

Kaveri also looks to have decent yields as there s few harvested offerings, actualy only two that are significant, the 7400K and the 860K.

As for Beema yields are inherently good due to the minimalist size, besides i think that there s more built in redundancies in the 28nm chips, particularly for Kaveri.

 

[cbn]

Vishera 8C should have good yields given the process maturity

Over 90% at this point? Would even mid to high 90% be realistic? Or is Vishera still too big for that kind of yield?

 

[Abwx]

Over 90% at this point? Would even mid to high 90% be realistic? Or is Vishera still too big for that kind of yield?

Thats too high i think, they should have yields comparable to a mature 28nm process for 200mm2 chips, that is about 50%, the remaining 50% will be used mainly as 6C offerings with the 4C being marginal overall, only in this sense the yields could reach 80-90%, the main advantage of this CPU cost wise is the amortized process, surely that AMD makes the essential of their bread and butter with this CPU line, it s unfortunate that they didnt seems to have envisioned this two years ago...

 

[cbn]

Thats too high i think, they should have yields comparable to a mature 28nm process for 200mm2 chips, that is about 50%, the remaining 50% will be used mainly as 6C offerings with the 4C being marginal overall, only in this sense the yields could reach 80-90%

I was hoping the yield would be higher for the 8C.

However, it does make sense to me the 4C would be a rare item (what are the chances for two defects on the same die? surely this has to be on the low side)

P.S. Here are # Newegg reviews for FX-8350, FX-6300 and FX-4300:

http://www.newegg.com/Product/Produc...82E16819113284

(FX-8350 has 1845 reviews and was launched Oct 23,2012 according to cpu world)

http://www.newegg.com/Product/Produc...82E16819113286

(FX-6300 has 986 reviews and was launched Oct 23,2012 according to cpu world)

http://www.newegg.com/Product/Produc...82E16819113287

(FX-4300 has 141 reviews and was launched Oct 23, 2012 according to cpu world)

Surprisingly the FX-8350 has twice as many reviews as the FX-6300 and thirteen times as many as the FX-4300. This despite all three processors being launched on the same day and thus being available on the market for the same amount of time.

That is a lot of reviews for the FX-8350. Part of me wonders how much the number of reviews correlates to Newegg sales numbers? Or even more remotely AMD sales numbers? But it could also be that folks who paid the high purchase price for FX-8350 were more likely to leave a review than someone who bought FX-6300 or FX-4300.

 

[krumme]

The families that control mubadala and the oil wealth are not worried about themselves and their present day wealth. They are worried about what their families will be doing in 100+ years when they have nothing left but their accumulated warchest and nothing left to create jobs in their country outside of hoping for a sand exporting business.

In many ways they are a lot more forward looking than many western nations who are doing little more than creating a large debt for their future generations with little concern for seeding industries that will create stable long-lasting job growth.

We all can't become Uber drivers. Somebody has to be willing and able to create the products that the rest of plan to kick back and consume while living off our pensions or retirement savings

Agree - about 80% in this planet is uber drivers as of today. Everyone can be a small farmer in easter china or a factory worker in Thailand. The competition for the job is 100% - as uber drivers. There is no time, ressource and personal energy left for the slightest innovation. 14nm finfet is farther away than the moon. No wonder if those people driving their bicycle living day by day, using choal as heating is for fracking and frown of prius drivers

But what happens when the owners of future innovative centers - monopoly like companies as gf - add absolutely nothing but money and no competences? what is the direction then ??

 

[witeken]

Just out of curiosity, what would AMD's expected yields be for 32nm and 28nm at GF today?

32nm for Vishera 315mm2 die size
28nm for Kaveri 245mm2 die size
28nm for Beema/Mullins ~101mm2 die size

Percentage wise does anyone want to wager a guess for each of those different chips I've listed above?

Over 90% at this point? Would even mid to high 90% be realistic? Or is Vishera still too big for that kind of yield?

Lol, no. Yields will at most be 83% due to big rectangular die made on round wafer:

Dies per wafer =

(Pi x radius wafer²) / die area .. minus .. (Pi x wafer diameter) / sqrt(2 x die area)

(Pi x 150² / 315) - (Pi x 300 / sqrt(2 x 315) = 224.3994753 - something = 186.8502611.

Divide the second number by the first one = 0.8327 = 83.27%.

Now, how about yields?

Die yield = wafer yield x 1 / (1 + defects per unit area x die area)^N
Let's be optimistic.. (apparently N is 11.5-15.5 for a 40mm proces, and defects per cm² is 0.016-0.057)

Die yield = 0.90 x 1/(1 + 0.016 x 315)^11.5 = 0.51128 = 51.13%
186.85 x 0.5112 = 95.52 ~= 95

So your total yield from the whole 350mm wafer of silicon will be 42.57%; from the usable portion of the die it will be 51.12%. Maybe 28nm is mature and the defect density is now lower, but'll still be far from 90%.

 

[Idontcare]

Surprisingly the FX-8350 has twice as many reviews as the FX-6300 and thirteen times as many as the FX-4300. This despite all three processors being launched on the same day and thus being available on the market for the same amount of time.

That is a lot of reviews for the FX-8350. Part of me wonders how much the number of reviews correlates to Newegg sales numbers? Or even more remotely AMD sales numbers? But it could also be that folks who paid the high purchase price for FX-8350 were more likely to leave a review than someone who bought FX-6300 or FX-4300.

Yeah I wouldn't read anything into the relative weightings of newegg reviews by SKU. Too much personal psychology of the end-user/reviewer is tied up in their personal motivations for posting a review of their system (be it pride, ax to grind, etc.).

 

[Enigmoid]

Lol, no. Yields will at most be 83% due to big rectangular die made on round wafer:

Dies per wafer =

(Pi x radius wafer²) / die area .. minus .. (Pi x wafer diameter) / sqrt(2 x die area)

(Pi x 150² / 315) - (Pi x 300 / sqrt(2 x 315) = 224.3994753 - something = 186.8502611.

Divide the second number by the first one = 0.8327 = 83.27%.

Now, how about yields?

Die yield = wafer yield x 1 / (1 + defects per unit area x die area)^N
Let's be optimistic.. (apparently N is 11.5-15.5 for a 40mm proces, and defects per cm² is 0.016-0.057)

Die yield = 0.90 x 1/(1 + 0.016 x 315)^11.5 = 0.51128 = 51.13%
186.85 x 0.5112 = 95.52 ~= 95

So your total yield from the whole 350mm wafer of silicon will be 42.57%; from the usable portion of the die it will be 51.12%. Maybe 28nm is mature and the defect density is now lower, but'll still be far from 90%.

But that gives the number of 'perfect dies' presumably some of those are binned at 4 and 6 core variants. Also parts of the CPU such as cache usually have built in excess so that a small defect doesn't destroy the chip.

 

[Idontcare]

Lol, no. Yields will at most be 83% due to big rectangular die made on round wafer:

Dies per wafer =

(Pi x radius wafer²) / die area .. minus .. (Pi x wafer diameter) / sqrt(2 x die area)

(Pi x 150² / 315) - (Pi x 300 / sqrt(2 x 315) = 224.3994753 - something = 186.8502611.

Divide the second number by the first one = 0.8327 = 83.27%.

I suspect this portion of your post will confuse folks, so just to clarify for them I will add that what you are computing above is the "areal yield" which is not to be confused with die yield.

Areal yield is a little-used metric, and generally speaking anytime you see a yield number reported it will be the "die yield" number unless otherwise explicitly stated by the person or source reporting the number.

In practice we don't worry or discuss areal yield outside of the initial hypothetical layout and die placement considerations. (silicon budgets for design teams) At the operational level of the fab production the areal yield is irrelevant since it doesn't change (it is a fixed value dependent on the IC layout itself, not dependent on the operation status of the fab).

For large die, >400mm^2, where areal yield loss is a significant cost-adder to the per-die manufacturing cost it is common to see the yield numbers converted to absolute numbers instead of percentages. The absolute numbers are referred to as NUBs per wafer (Net Units Built per wafer).

This is done for practical considerations. If you are yielding 1 NUB, the percentage yield will be meaningless. And if you increase the yield from 1 NUB to 2 NUBs then the yield percentage will remain meaningless, but doubling the yield to 2 NUBs is not meaningless (you can halve your wafer starts at that point).

Now, how about yields?

Die yield = wafer yield x 1 / (1 + defects per unit area x die area)^N
Let's be optimistic.. (apparently N is 11.5-15.5 for a 40mm proces, and defects per cm² is 0.016-0.057)

Die yield = 0.90 x 1/(1 + 0.016 x 315)^11.5 = 0.51128 = 51.13%
186.85 x 0.5112 = 95.52 ~= 95

So your total yield from the whole 350mm wafer of silicon will be 42.57%; from the usable portion of the die it will be 51.12%. Maybe 28nm is mature and the defect density is now lower, but'll still be far from 90%.

Yield, the kind that people are referring to when they generically reference yield, will be the product of functional yield (Fy) and parametric yield (Py).

Yield = Fy x Py

If you have 90% functional yield and 60% parametric yield then you can expect 54% yield.

Of course there are other types of yield, but they will always be explicitly labeled with the appropriate adjective (test yield, package yield, shipping yield, etc.).

Functional yield is the part is improved with die harvesting. Parametric yield can be improved with die harvesting but only to the extent that your underlying process has such horrific within-die process variability as to create the opportunity for improving parametric yields by fusing of electrically funky parts of the die (a core that cannot clock over 800MHz without burning through 300W of power, for example).

 

[witeken]

Great post, Idontcare. I guess the point is that larger dies have a worse yield, so 90% for >300mm² is unlikely, certainly not when you add the lost areal portions of the wafer in.

 

[cbn]

Yield, the kind that people are referring to when they generically reference yield, will be the product of functional yield (Fy) and parametric yield (Py).

Yield = Fy x Py

If you have 90% functional yield and 60% parametric yield then you can expect 54% yield.

Of course there are other types of yield, but they will always be explicitly labeled with the appropriate adjective (test yield, package yield, shipping yield, etc.).

Functional yield is the part is improved with die harvesting. Parametric yield can be improved with die harvesting but only to the extent that your underlying process has such horrific within-die process variability as to create the opportunity for improving parametric yields by fusing of electrically funky parts of the die (a core that cannot clock over 800MHz without burning through 300W of power, for example).

Thanks for the formula and the clear explanation on how that works.

 

[cbn]

Great post, Idontcare. I guess the point is that larger dies have a worse yield, so 90% for >300mm² is unlikely

Going by the definition I just read I think it would be technically possible to have very high yields even with a very large die, but it would have to be accompanied by higher amounts of die harvesting, etc

Of course, all this makes talk of yields somewhat more confusing at the same time because the definition of what makes a chip can be different depending on the point of view.

 

[Abwx]

P.S. Here are # Newegg reviews for FX-8350, FX-6300 and FX-4300:


That is a lot of reviews for the FX-8350. Part of me wonders how much the number of reviews correlates to Newegg sales numbers? Or even more remotely AMD sales numbers? But it could also be that folks who paid the high purchase price for FX-8350 were more likely to leave a review than someone who bought FX-6300 or FX-4300.

Even if you adjust drasticaly the numbers to compensate from some theorical bias you ll get the same picture, here the 8350 outsold all other models including the cheaper 8320, the perf/price ratio of this CPU is very favourable, think about it, 100% and 33% more perfs than the 4C/6C respectively and then look at the prices deltas.

 

[cbn]

Deleted.

 

[Idontcare]

Great post, Idontcare. I guess the point is that larger dies have a worse yield, so 90% for >300mm² is unlikely, certainly not when you add the lost areal portions of the wafer in.

Going by the definition I just read I think it would be technically possible to have very high yields even with a very large die, but it would have to be accompanied by higher amounts of die harvesting, etc

Of course, all this makes talk of yields somewhat more confusing at the same time because the definition of what makes a chip can be different depending on the point of view.

90% yields are possible, but it takes a lot of "design for manufacturing" effort (read: $$$) in the layout phase of the chip's design.

Lots of redundancy must be built in, so your 300mm^2 chip might only have 250mm^2 of active silicon and the remainder is just dead space built into the chip to enable die-harvesting.

If you are looking for 90% yield on a 300mm2 chip without resorting to redundancy then it is pretty much impossible.

There are two sides to the discussion though, because functional yield is itself essentially a production cost variable (lower yield means higher production cost per sellable die, higher yield only means lower production costs per die) whereas parametric yield is itself essentially an ASP or revenue variable (better speedbins, better power profiles, enabling higher MSRP, etc.).

 

[TurtleCrusher]

but you don't seem to understand what general public means. The general tech public knows little more than the fact that a lower nm name is better, and since the names are the same, Intel obviously can't be much ahead. .

The general public, let alone most people who own a computer, have no idea what's really in the case. They also have very little inclination to as well since they're actual users and not shills/investors like we see in forums like this.

Does it work? Is it fast? Those are the two questions most people, even in tech, care about. The second one is subjective an can be clouded with marketing for "perceived" fast.