How long does Intel produce prior-gen parts in their fabs?

[aigomorla]

i heard its somewhere between 5 yrs before they can migrate new node processes onto the old fab.

Also depends on how large the demand for old nodes are.
If there is still a high demand intel will delay the migration.
If new node processors is in higher demand, they will retire the old one out faster.

But 5 yrs is what i was told.
Thats about the time required to fill all the late RMA's b4 the cpu warranty is completely extinguished.

I wonder what they did with all those flawed Haswell Processors?

Intel is NOT AMD.
When you have access to primary fabs all bowing down to you, you dont need to worry about waste.
There has not been a dual core intel cpu which had 2 cores locked cuz 1 was faulty, or the cpu was faulty.
Also intel's fabs are highly specialized so they have less waste then other FABs.

Its not TSMC where there pumping dies for serveral companies.

And finally yes they do check every last one of them.
They put the finished cpu's in a machine which runs tests on it and tells which one is faulty.
That one is typically sent to RnD for inspection on how and why it failed during the fab processor.

Intel also loves failed cpu's if i recall.
My sponsor would get brownie points with intel if he ever had a failed cpu.
They really love finding out why something broke, and how it broke.
(my sponsor told me its probably because their bored as hell... how many intel cpu's do u honestly know which failed of natural reasons? NO OCing)

its an older process design but they do not convert all their factories over to 22nm overnight and the tooling and such they will use for a while to make chipsets and such.

i mean they are still shipping 32nm cpus, like the celeron 847 and such on all those itx boards.

New chips built on a new manufacturing process require an entirely new fab. The oldest fabs get stuck manufacturing the chipsets and obviously ever decreasing quantities of old chips.

I think the oldest designs, the decades old ones, are possibly contract manufactured at some archaic fabs around the world, possibly.

Add to these two...
Intel can not swap nodes overnight.
The process takes a VERY LONG TIME.
Sometimes its easier just to build a new fab then migrate an old one to a new process node.

 

[RU482]

different skus have different life spans.

Standard skus have ~3yr life spans, Embedded skus have ~7yr life spans.
You can guarantee if they have a customer that justifies the cost, they'll keep producing a CPU

 

[Plimogz]

Intel is NOT AMD.

No, they're not. But Intel has for a long time segmented their products based on cache size, and in doing so have in all likelihood been binning and selling "failed chips" just as AMD has done. "Cores" make up a relatively small part of overall die size, whereas the cache area is pretty big. So it makes sense that if imperfections crop up in a cache area, they would just partition the faulty area off and sell it off as a lower-end SKU.

This 2500K of mine, for instance. We all notice when the HT is turned off, but that 2MB of missing L3? You can't tell me that this is necessarily the product of pure segmentation. I find way more reasonable to think that 2500K's L3 specs are the product of managing the risk of some minor defects popping up in the wafers by releasing a SKU which by definition can allow for a given margin of manufacturing error.

To say that there isn't a junk pile behind Intel fabs onto which dies -- which can't be sold because they don't adhere to even the lowest SKUs -- get dumped is just wishful thinking. For it to be true would mean that they could just forgo testing the CPUs altogether, because all would be equally good; with every single one being just as good as any other.

Though I'll grant you that the Intel refuse piles might be pretty small, considering not only their industry leading know-how, but also their currently advantageous market position which enables them to profitably market chips which don't at all push the limit of what they could actually do, if they really had do.

 

[Insert_Nickname]

This 2500K of mine, for instance. We all notice when the HT is turned off, but that 2MB of missing L3? You can't tell me that this is necessarily the product of pure segmentation. I find way more reasonable to think that 2500K's L3 specs are the product of managing the risk of some minor defects popping up in the wafers by releasing a SKU which by definition can allow for a given margin of manufacturing error.

Spot on. There are only 4 different dies used for all Ivy Bridge CPUs (unless you count the LGA-2011 ones). If you look closely enough at the product stack, you can even tell where a "defective" die ends up...

 

[Topweasel]

Spot on. There are only 4 different dies used for all Ivy Bridge CPUs (unless you count the LGA-2011 ones). If you look closely enough at the product stack, you can even tell where a "defective" die ends up...

Another example is the SB 2011 chips. They are all (not counting the 38xx) 8 cores with 2 cores and a bunch of cache disabled server chips. With the 3930 having more cache disabled than the 3960. They say they did that for the power envelope. But reality is in the middle they made those specs and sku's so that they could bin what would probably be problematic chips so that they wouldn't fail tests.

They make outlandish statements like "we disable our chips just for different market segments" and "we wouldn't ever sell a chip that failed any testing" is just to protect themselves from the backlash of the Pentium bug. They also say it because they can. They know their dies pretty well. They know what the manufacturing characteristics of their wafers are like. They can tell right off the bat before QCing the chips, which ones would have a hard time passing a test and why. The only thing they do differently than AMD is bin them before testing instead of later. Allowing them to say "they never failed a test".

 

[Wall Street]

I believe that Intel has indicated in the past they they actually do make Core i5 dies. They aren't i7 dies with 2 MB cache and HT disabled, they are actually 6 MB dies that never had hyperthreading.

 

[nehalem256]

I believe that Intel has indicated in the past they they actually do make Core i5 dies. They aren't i7 dies with 2 MB cache and HT disabled, they are actually 6 MB dies that never had hyperthreading.

Given how tightly coupled hyperthreading is to the core there is pretty much no way to make an Ivy Bridge without the hyperthreading parts.

And if you look at a die shot of Ivy/Sandy Bridge it doesn't look like there is any way that cutting out 2MB of L3 cache would even work to reduce die size and still produce a rectangular die

 

[ShintaiDK]

I believe that Intel has indicated in the past they they actually do make Core i5 dies. They aren't i7 dies with 2 MB cache and HT disabled, they are actually 6 MB dies that never had hyperthreading.

Thats not correct. Its simply fused off.

 

[Duvie]

They have multiple fabs and not all of them upgrade process at same time, so it is likely there are some fabs still producing sandy bridge (close to the end I think), ivy bridge and newer Haswell chips.

 

[R0H1T]

They have multiple fabs and not all of them upgrade process at same time, so it is likely there are some fabs still producing sandy bridge (close to the end I think), ivy bridge and newer Haswell chips.

So does that mean over a period of time, say 5~10yrs, all of their fabs will be upgraded eventually ?

 

[ShintaiDK]

So does that mean over a period of time, say 5~10yrs, all of their fabs will be upgraded eventually ?

65nm is being phased out now with the chipsets moving from 65nm to 32nm. Unlike TSMC, etc that still got huge amounts of 150nm and 180nm. Intel is in a much more rapid change due to the benefits of its own products. 65nm funny enough latested longer than 45nm.

 

[R0H1T]

65nm is being phased out now with the chipsets moving from 65nm to 32nm. Unlike TSMC, etc that still got huge amounts of 150nm and 180nm. Intel is in a much more rapid change due to the benefits of its own products. 65nm funny enough latested longer than 45nm.

To be more precise what I meant was that all of their production lines/fabs will eventually be upgraded to say 32nm over a period of time, except for embedded & some other ancient parts, right ?

 

[ShintaiDK]

To be more precise what I meant was that all of their production lines/fabs will eventually be upgraded to say 32nm over a period of time, except for embedded & some other ancient parts, right ?

Not upgraded to 32nm. 32nm capacity is shrinking(Or at best stagnant) in favour of 14nm.

 

[R0H1T]

Not upgraded to 32nm. 32nm capacity is shrinking(Or at best stagnant) in favour of 14nm.

Umm yeah that'd be at least 32nm, for chipsets mainly, anyways I wonder how long does it take for the given transition :hmm:

 

[ShintaiDK]

D1XM1+DX1M2 and fab42 is entirely new 14nm fabs with 450mm wafer support. Fab24 will get retrofitted for 14nm.

I think most old (intel)fabs will be demolished in the coming time due to lack of 450mm capability.

 

[Sheep221]

D1XM1+DX1M2 and fab42 is entirely new 14nm fabs with 450mm wafer support. Fab24 will get retrofitted for 14nm.

I think most old (intel)fabs will be demolished in the coming time due to lack of 450mm capability.

Demolishing the fab doesn't make any sense. The more dense manufacturing process does not mean the entire building is changed or new one is need to be built, in fact lithographic steppers and ion implantation systems are the only machines that get upgraded, probably something else ofc as well, but the entire AC, heating, transportation, cleaning, power and all other systems within the building remain the same, because they don't correlate with smaller process. So they either gonna increase the capacity for newest chips and upgrade all fabs, or they will remain like they are and the production of present chips will be moved there.

To answer the first question, the CPUs remain in production few more years after they were called end of life. However the sandy bridge is going to be good for long here. And they are in great overstock, so it will be possible to buy these CPUs used in probably next 10 years.

i heard its somewhere between 5 yrs before they can migrate new node processes onto the old fab.

Also depends on how large the demand for old nodes are.
If there is still a high demand intel will delay the migration.
If new node processors is in higher demand, they will retire the old one out faster.

But 5 yrs is what i was told.
Thats about the time required to fill all the late RMA's b4 the cpu warranty is completely extinguished.

Intel is NOT AMD.
When you have access to primary fabs all bowing down to you, you dont need to worry about waste.
There has not been a dual core intel cpu which had 2 cores locked cuz 1 was faulty, or the cpu was faulty.
Also intel's fabs are highly specialized so they have less waste then other FABs.

Its not TSMC where there pumping dies for serveral companies.

And finally yes they do check every last one of them.
They put the finished cpu's in a machine which runs tests on it and tells which one is faulty.
That one is typically sent to RnD for inspection on how and why it failed during the fab processor.

Intel also loves failed cpu's if i recall.
My sponsor would get brownie points with intel if he ever had a failed cpu.
They really love finding out why something broke, and how it broke.
(my sponsor told me its probably because their bored as hell... how many intel cpu's do u honestly know which failed of natural reasons? NO OCing)

Add to these two...
Intel can not swap nodes overnight.
The process takes a VERY LONG TIME.
Sometimes its easier just to build a new fab then migrate an old one to a new process node.

The thing is that Intel does the same thing but the cores get disabled for permanent to prevent their unlocking. Celerons and Pentiums are faulty i7s or i3s respectively, with the defective cores and cache. There is also full series of CPUs with defective graphics cores, such as i5-2450P and i5-2550K. I don't believe that either intel or AMD ever throws away any partially defective wafers. There are probably only 2-3 different nodes or so in production, all the other CPUs are only intentionally slowed down, have features removed or are made from defective wafers.

I believe that Intel has indicated in the past they they actually do make Core i5 dies. They aren't i7 dies with 2 MB cache and HT disabled, they are actually 6 MB dies that never had hyperthreading.

i5s are made from i7 dies, but the HT is removed and cache is lowered

 

[24601]

Demolishing the fab doesn't make any sense. The more dense manufacturing process does not mean the entire building is changed or new one is need to be built, in fact lithographic steppers and ion implantation systems are the only machines that get upgraded, probably something else ofc as well, but the entire AC, heating, transportation, cleaning, power and all other systems within the building remain the same, because they don't correlate with smaller process. So they either gonna increase the capacity for newest chips and upgrade all fabs, or they will remain like they are and the production of present chips will be moved there.

To answer the first question, the CPUs remain in production few more years after they were called end of life. However the sandy bridge is going to be good for long here. And they are in great overstock, so it will be possible to buy these CPUs used in probably next 10 years.

Sometimes you can get local governments to pay for most of the construction cost of a Fab just for the political claims the politician(s) can make afterwards.

 

[ShintaiDK]

Demolishing the fab doesn't make any sense. The more dense manufacturing process does not mean the entire building is changed or new one is need to be built, in fact lithographic steppers and ion implantation systems are the only machines that get upgraded, probably something else ofc as well, but the entire AC, heating, transportation, cleaning, power and all other systems within the building remain the same, because they don't correlate with smaller process. So they either gonna increase the capacity for newest chips and upgrade all fabs, or they will remain like they are and the production of present chips will be moved there.

To answer the first question, the CPUs remain in production few more years after they were called end of life. However the sandy bridge is going to be good for long here. And they are in great overstock, so it will be possible to buy these CPUs used in probably next 10 years.
The thing is that Intel does the same thing but the cores get disabled for permanent to prevent their unlocking. Celerons and Pentiums are faulty i5s and i7s respectively, with the defective cores and cache. There is also full series of CPUs with defective graphics cores, such as i5-2450P and i5-2550K. I don't believe that either intel or AMD ever throws away any partially defective wafers. There are probably only 2-3 different nodes or so in production, all the other CPUs are only intentionally slowed down, have features removed or are made from defective wafers.i5s are made from i7 dies, but the HT is removed and cache is lowered

Celerons and Pentiums are not faulty i5 or i7. Intel only disable cores on server/workstation class chips. But rarely for being faulty.

And Intel sold/demolished fabs before. Most fabs simply cant be retrofitted for 450mm. And thats what the future will demand for top profits. Also why 700 people got laid off on an older fab in Oregon. And 2 new fabs constructed instead.

 

[VirtualLarry]

Celerons and Pentiums are not faulty i5 or i7. Intel only disable cores on server/workstation class chips. But rarely for being faulty.

It really wouldn't surprise me at all to find out that some Celerons, are actually made from defective i7s. It just doesn't make economic sense to throw them out, when they can laser off or fuse off the defective cores, and most of the cache, and sell them as a low-end chip, like the single-core Celeron with HyperThreading. When you CAN do that, why wouldn't you? It's not like Intel likes to throw away something that could make them money. What, because they have pride in mfg, and don't want to admit that they recycle silicon?

And Intel sold/demolished fabs before. Most fabs simply cant be retrofitted for 450mm. And thats what the future will demand for top profits. Also why 700 people got laid off on an older fab in Oregon. And 2 new fabs constructed instead.

I can understand selling an older fab, perhaps, but likewise, I just cannot see them outright demolishing it, considering it's worth money. Why do you think Intel is such a rich corporation, that they can just waste money?

 

[ShintaiDK]

It really wouldn't surprise me at all to find out that some Celerons, are actually made from defective i7s. It just doesn't make economic sense to throw them out, when they can laser off or fuse off the defective cores, and most of the cache, and sell them as a low-end chip, like the single-core Celeron with HyperThreading. When you CAN do that, why wouldn't you? It's not like Intel likes to throw away something that could make them money. What, because they have pride in mfg, and don't want to admit that they recycle silicon?

I can understand selling an older fab, perhaps, but likewise, I just cannot see them outright demolishing it, considering it's worth money. Why do you think Intel is such a rich corporation, that they can just waste money?

Fabs are mainly subsidized. And remember its just a concrete shell. Its not there the big cost is.

And yet again, there are no faulty i5s and i7s that goes as Celerons and Pentiums. To apply your own logic, Intel should sell tricores too. Yet they do not.

 

[VirtualLarry]

Fabs are mainly subsidized. And remember its just a concrete shell. Its not there the big cost is.

It's a lot more than "a concrete shell". They have to install supports down into the bedrock for the litho equipment. It's quite involved.

And yet again, there are no faulty i5s and i7s that goes as Celerons and Pentiums. To apply your own logic, Intel should sell tricores too. Yet they do not.

Well, I don't see how that follows from my logic. If Intel sold tricores, that might bring down their ASP, because if more people purchased tri-cores, then Intel would have to cripple working quad-cores that could have been sold as quad-cores.

Conversely, if I used your logic (that Intel doesn't cripple / re-use chips), then where did that single-core Sandy Bridge celeron with HyperThreading come from?

Clearly, they do not produce a single-core die, so it MUST be made from crippled chips.

My theory is that crippled quad-cores also get sent to that pile. Else, where would they go? As you said, Intel doesn't sell tri-cores.

But some of their lowest-end SKUs could easily be a "dumping grounds" for defective dies.

 

[Magic Carpet]

65nm funny enough latested longer than 45nm.

Strangely, Intel didn't use 45/32nm for IB chipsets. Wasn't it one generation behind?

 

[ShintaiDK]

Strangely, Intel didn't use 45/32nm for IB chipsets. Wasn't it one generation behind?

It was essentially the same previously as well. 65nm and (some)45nm CPUs using 130nm chipsets.

5, 6 and 7 series chipsets is 65nm. 8 and 9 series will be 32nm.

 

[VirtualLarry]

It was essentially the same previously as well. 65nm and (some)45nm CPUs using 130nm chipsets.

5, 6 and 7 series chipsets is 65nm. 8 and 9 series will be 32nm.

It seems amazing to me, that with such large process disparities between the CPU nm and the chipset nm, that they would be able to work together with electrically-compatible buses and signalling.

 

[Sheep221]

Sometimes you can get local governments to pay for most of the construction cost of a Fab just for the political claims the politician(s) can make afterwards.

I'm not sure that they afford to do that anytime soon, western world economy is about to bankrupt in no time.

Fabs are mainly subsidized. And remember its just a concrete shell. Its not there the big cost is.

And yet again, there are no faulty i5s and i7s that goes as Celerons and Pentiums. To apply your own logic, Intel should sell tricores too. Yet they do not.

Or they would just disable the other core as well and sell it as dual core. I don't know why Intel never sold tri cores but it might be a design or corporate/marketing decision to just sell duals and quads or to not expose they are doing any activity of this kind.
On other hands, it's very unlikely that intel would have separate nodes for every type and probably even more processor model.
Within sandy bridge architecture, there are total 123 CPUs, 51 desktop and 72 mobile within 10 model ranges, but there are only 3 die sizes the 131mm², 149mm² and 216mm². Which means they are making dual cores natively but probably only highest CPU for dual core and quad core line and the rest are defective and downgraded ones. Only desktop i3 and i7 are being made natively, the i3 is on both 131 and 149 and i7 on 216mm² die.
The desktop i5 and mobile quad core i7 are made from desktop i7 node, the mobile i5, mobile dual core i7, mobile i3, i5-2390T and both desktop and mobile celerons and pentiums are made from both of desktop i3 nodes.
Which leaves us with i3-2125, i3-2130 and i7-2700K are being only CPUs made natively. The rest is derived from them, either by downgrading their specs, removing features or disabling cores and IGP. I am aware that official numbers don't prove that quad cores are sold as disabled dual cores, but there are for sure dual cores with i7 die out there, but we never find that out as sandys are not really good to delid and it would be physically impossible to check all of them around in an effort of proving this phenomenon.

It seems amazing to me, that with such large process disparities between the CPU nm and the chipset nm, that they would be able to work together with electrically-compatible buses and signalling.

Chipset are not as technically complex as CPUs and they don't need to be upgraded this often. Chipset is not doing much stuff nowadays either, it's just I/O controller.