This is a big change for Samsung as they'll have to have their fabs capable of handling >600 mm^2 designs and >5 Ghz frequencies. They did some high performance stuff like this in the past (DEC Alpha) but that is decades ago on 250 nm process. IBM could go the chiplet route (and they kind of do already on zSeries with the system controllers being on package with multiple CPU dies) to avoid the high size criteria but there is still the frequency aspect.
POWER10 could likely move to Samsung with little issue but zSeries is used in some critical US government facilities and moving manufacturing overseas may meet some resistance.
On a tangent, wasn't Nvidia's first plan for this gen Ampere @ Samsung 10nm? AFAIK, that fell through and Turing was quickly spun up on familiar TSMC 12FFN to fill in the gap. Turing showed up really late on the roadmaps, which is probably where that rumor comes from.
IBM also demoed 7nm EUV first with Samsung, several years ago. IBM & Nvidia probably have a vested interest (more than OpenPOWER) in getting large ASIC working on Samsung processes, since the "only option" is TSMC (Intel tends to shy away from making larger dies).
With 7nm processes widely expected to be a long term node for the industry, the way 14nm has been, I wouldn't be surprised if Samsung does announce a new FAB somewhere within the next 2 years. May not be 7nm, could be part of their 5nm ramp up as capacity shuffling happens
Well. There goes Global F(l)oundries down to the bin. Let us see how they spin this one. So after AMD spun off GF in 2005, there was only 1 node they made somewhat reasonably commerically viable for a small period 28nm bulk. They basically swallowed Chartered, IBM Micro and still had all the FD-SOI nodes in the world and still cannot do any s**t out of it. Clever IBM, hive off its Fabs to the universal graveyard and now move to some other competitor. With GF now deciding that 12nm is their current best 1-1.5 generations behind others, what clients will they have for their supposedly top FinFET and FD-SOI nodes??
IBM had to pay ~$1.5B USD to get GF to take over it's fabs. AFAIK, IBM's fabs were run at quite a loss. Why GF thought it was a good idea to take on IBMs fabs, remains a mystery.
Well as much as I can recall the 14nm costume should be a 14nm FinFET structure with SOI wafer, if the EUV turns to be cost cutting enough for additional benefits that it brings the migration will be swift and rather sooner than later. The 7nm EUV Samsung presented so far is based on High Density Lib, the road map also shows "5nm EUV" which will actually be a same o 7nm with Ultra High Density lib, but I doubt IBM will use that (because performance goals they have). Judging by announcements this should be a late 2019 shirt if IBM is serious & wants to gain some momentum.
So some bright spark exec at IBM decided that they'll cut costs by going fabless and letting GloFo pay for fabbing R&D.. then GloFo turns around and says "BTW we don't feel like paying for R&D anymore"... and now IBM has had to go, hat in hand, to Samsung to ask to use their fab. I wonder how much that's costing IBM... I wonder if they are prepared for the low-quality shitfest that is Samsung's power-leaking chips?
Looking back a year and a half later, we now know what it cost IBM: nearly everything. They embarrassingly lost all three of the exaflop supercomputer bids, even after submitting extra bids without Nvidia (or anyone else; who even thought that was a good idea? A non-accelerator POWER9 server is a laughably poor concept for something that is supposed to scale while consuming less power than a small city). IBM's backhanded desperation move probably convinced Nvidia to promote Arm to their main "non x86" platform, largely supplanting POWER. Nvidia isn't even a Platinum (top tier) OpenPOWER member, anymore, sliding down to the lowest tier (Mellanox, before acquisition, was Gold level - second tier). Nvidia was one of the few co-founders of OpenPOWER, alongside IBM and formerly held a top tier Platinum spot.
I think it's pretty clear, now. IBM's short term cost cutting measure has lost them their fabs, future supercomputer contracts (relevancy, as well), and their core partners. Who's left at the high level in OpenPOWER? A russian company, great. They also have Inspur. Well, recent news isn't rosy for that partner, either. Not to mention, IBM historically partnered with Lenovo for HW (another absence noted). Finally, Google, who has their own ambitions and has largely avoided using POWER ISA. I wonder if Google keeps paying the membership fee, since it is barely a rounding error for them and it garners them a bit more insight into IBM's supercomputer program. The next tier is even sadder. Lost two (Mellanox and Wistron), leaving just Hitachi, who has been out of relevance (in the HPC market) for a while.
IBM, in one boneheaded move, has managed to further sink their hardware into irrelevancy. Many (though not all) of the fancy services IBM runs depended on their big iron staying relevant. While they still make money nowadays, their moves are merely completing their slow drive into the abyss of giants past. Maybe I'm being melodramatic. But without the gateway (and anchor) of IBM hardware, the competitive market for services (what IBM is shifting to) is much harsher. Why buy Oracle Cloud if you aren't locked in via the tangled mess of Peoplesoft? Why buy IBM services if you aren't locked in to their Z/POWER servers? IBM would have to wholly reinvent themselves, and they haven't shown any inkling of that happening.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
24 Comments
Back to Article
Kevin G - Friday, December 21, 2018 - link
This is a big change for Samsung as they'll have to have their fabs capable of handling >600 mm^2 designs and >5 Ghz frequencies. They did some high performance stuff like this in the past (DEC Alpha) but that is decades ago on 250 nm process. IBM could go the chiplet route (and they kind of do already on zSeries with the system controllers being on package with multiple CPU dies) to avoid the high size criteria but there is still the frequency aspect.POWER10 could likely move to Samsung with little issue but zSeries is used in some critical US government facilities and moving manufacturing overseas may meet some resistance.
CoreDuo - Friday, December 21, 2018 - link
Samsung has a huge fab in Austin, TX.jeremyshaw - Friday, December 21, 2018 - link
14nm, iirc. The 10 and 7nm stuff is in Korea.On a tangent, wasn't Nvidia's first plan for this gen Ampere @ Samsung 10nm? AFAIK, that fell through and Turing was quickly spun up on familiar TSMC 12FFN to fill in the gap. Turing showed up really late on the roadmaps, which is probably where that rumor comes from.
IBM also demoed 7nm EUV first with Samsung, several years ago. IBM & Nvidia probably have a vested interest (more than OpenPOWER) in getting large ASIC working on Samsung processes, since the "only option" is TSMC (Intel tends to shy away from making larger dies).
CoreDuo - Friday, December 21, 2018 - link
Doesn't mean that it has to stay that way. If push comes to shove, they could retrofit or expand it.FullmetalTitan - Saturday, December 22, 2018 - link
With 7nm processes widely expected to be a long term node for the industry, the way 14nm has been, I wouldn't be surprised if Samsung does announce a new FAB somewhere within the next 2 years. May not be 7nm, could be part of their 5nm ramp up as capacity shuffling happensrocketbuddha - Friday, December 21, 2018 - link
Well. There goes Global F(l)oundries down to the bin. Let us see how they spin this one. So after AMD spun off GF in 2005, there was only 1 node they made somewhat reasonably commerically viable for a small period 28nm bulk.They basically swallowed Chartered, IBM Micro and still had all the FD-SOI nodes in the world and still cannot do any s**t out of it.
Clever IBM, hive off its Fabs to the universal graveyard and now move to some other competitor. With GF now deciding that 12nm is their current best 1-1.5 generations behind others, what clients will they have for their supposedly top FinFET and FD-SOI nodes??
LOL!
jeremyshaw - Friday, December 21, 2018 - link
IBM had to pay ~$1.5B USD to get GF to take over it's fabs. AFAIK, IBM's fabs were run at quite a loss. Why GF thought it was a good idea to take on IBMs fabs, remains a mystery.ZeDestructor - Sunday, December 23, 2018 - link
The answer to that is obvious: the figured increased capacity would get them more contracts and revenue. Unfortunately for em, it didn't pan out.ZolaIII - Friday, December 21, 2018 - link
Well as much as I can recall the 14nm costume should be a 14nm FinFET structure with SOI wafer, if the EUV turns to be cost cutting enough for additional benefits that it brings the migration will be swift and rather sooner than later. The 7nm EUV Samsung presented so far is based on High Density Lib, the road map also shows "5nm EUV" which will actually be a same o 7nm with Ultra High Density lib, but I doubt IBM will use that (because performance goals they have). Judging by announcements this should be a late 2019 shirt if IBM is serious & wants to gain some momentum.The_Assimilator - Friday, December 21, 2018 - link
So some bright spark exec at IBM decided that they'll cut costs by going fabless and letting GloFo pay for fabbing R&D.. then GloFo turns around and says "BTW we don't feel like paying for R&D anymore"... and now IBM has had to go, hat in hand, to Samsung to ask to use their fab. I wonder how much that's costing IBM... I wonder if they are prepared for the low-quality shitfest that is Samsung's power-leaking chips?jeremyshaw - Friday, July 3, 2020 - link
Looking back a year and a half later, we now know what it cost IBM: nearly everything. They embarrassingly lost all three of the exaflop supercomputer bids, even after submitting extra bids without Nvidia (or anyone else; who even thought that was a good idea? A non-accelerator POWER9 server is a laughably poor concept for something that is supposed to scale while consuming less power than a small city). IBM's backhanded desperation move probably convinced Nvidia to promote Arm to their main "non x86" platform, largely supplanting POWER. Nvidia isn't even a Platinum (top tier) OpenPOWER member, anymore, sliding down to the lowest tier (Mellanox, before acquisition, was Gold level - second tier). Nvidia was one of the few co-founders of OpenPOWER, alongside IBM and formerly held a top tier Platinum spot.I think it's pretty clear, now. IBM's short term cost cutting measure has lost them their fabs, future supercomputer contracts (relevancy, as well), and their core partners. Who's left at the high level in OpenPOWER? A russian company, great. They also have Inspur. Well, recent news isn't rosy for that partner, either. Not to mention, IBM historically partnered with Lenovo for HW (another absence noted). Finally, Google, who has their own ambitions and has largely avoided using POWER ISA. I wonder if Google keeps paying the membership fee, since it is barely a rounding error for them and it garners them a bit more insight into IBM's supercomputer program. The next tier is even sadder. Lost two (Mellanox and Wistron), leaving just Hitachi, who has been out of relevance (in the HPC market) for a while.
IBM, in one boneheaded move, has managed to further sink their hardware into irrelevancy. Many (though not all) of the fancy services IBM runs depended on their big iron staying relevant. While they still make money nowadays, their moves are merely completing their slow drive into the abyss of giants past. Maybe I'm being melodramatic. But without the gateway (and anchor) of IBM hardware, the competitive market for services (what IBM is shifting to) is much harsher. Why buy Oracle Cloud if you aren't locked in via the tangled mess of Peoplesoft? Why buy IBM services if you aren't locked in to their Z/POWER servers? IBM would have to wholly reinvent themselves, and they haven't shown any inkling of that happening.