Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
And then there were three: Global Foundries announces its intention to abandon the leading edge
https://www.anandtech.com/show/13277/globalfoundries-stops-all-7nm-development
For those not familiar with it, there are four foundries in the world that are competitive in offering the best process nodes for chips such as CPUs and GPUs: Intel, Samsung, TSMC, and Global Foundries. Global Foundries got its start when Abu Dhabi bought AMD's fabs, and later IBM's fabs and some others. They're now officially giving up, and leaving the cutting edge to Intel, Samsung, and TSMC alone.
Part of the contract of Global Foundries buying AMD's fabs was that AMD had to continue to fabricate chips there, while Global Foundries had to continue to push forward with advanced process nodes. Presumably they've cut a deal to allow AMD to fabricate its chips wherever they please, which will probably mean TSMC. The wafer share agreement meant that AMD had to buy at least some amount of wafers from Global Foundries, though they could also buy wafers from other fabs if they so desired.
For many years, AMD and Nvidia could use the same fabs, and generally used TSMC. For the 16/14/12 nm generation, Nvidia stayed at TSMC, while AMD moved to Global Foundries. That didn't work out well for AMD on the GPU side of things, though it's unclear whether Global Foundries was part of the problem. Presumably in the future, they'll both be back at TSMC, and any differences between AMD's and Nvidia's products will not be due to process node differences.
For those not familiar with it, there are four foundries in the world that are competitive in offering the best process nodes for chips such as CPUs and GPUs: Intel, Samsung, TSMC, and Global Foundries. Global Foundries got its start when Abu Dhabi bought AMD's fabs, and later IBM's fabs and some others. They're now officially giving up, and leaving the cutting edge to Intel, Samsung, and TSMC alone.
Part of the contract of Global Foundries buying AMD's fabs was that AMD had to continue to fabricate chips there, while Global Foundries had to continue to push forward with advanced process nodes. Presumably they've cut a deal to allow AMD to fabricate its chips wherever they please, which will probably mean TSMC. The wafer share agreement meant that AMD had to buy at least some amount of wafers from Global Foundries, though they could also buy wafers from other fabs if they so desired.
For many years, AMD and Nvidia could use the same fabs, and generally used TSMC. For the 16/14/12 nm generation, Nvidia stayed at TSMC, while AMD moved to Global Foundries. That didn't work out well for AMD on the GPU side of things, though it's unclear whether Global Foundries was part of the problem. Presumably in the future, they'll both be back at TSMC, and any differences between AMD's and Nvidia's products will not be due to process node differences.
Comments
Anything remotely cutting edge is always over priced and at Best Buy they ALWAYS want to sell you insurance to make sure the items/peripherals,products actually work which i find to be an appalling business ordeal.
So no matter what is happening behind closed doors,all i see in the biggest store is over priced stuff and imo a rip off because they do not honor any kind of support unless you pay EXTRA.
Never forget 3 mile Island and never trust a government official or company spokesman.
They are not going to build a 7nm plant so that basically leaves licencing or selling what they have developed. If so the question would be which companies might be interested.
Would any of Intel, Samsung or TSMC be interested? Not TSMC and I am not sure Samsung or Intel would be either - for all Intel's issues.
SMIC? Like GF they also partner with Imec; Qualcomm would - presumably - be supportive; and they are c. 2 generations behind. And anyone interested will require deep financial pockets a box they probably check as well.
Edit: I suppose AMD would, in theory, also be an option but a) they need 7nm "now" and b) they haven't got the financial pockets.
Also entirely possible that they may have had to reach agreement with AMD as you suggest (doubt either party would want to go to court).
As far as AMD buying wafers of course its the future 7nm wafers that are the issue. There was / is probably something in the agreement about them buying e.g. so many 14nm. Since GF are only giving up on 7nm this can continue. Maybe the price has changed!
There are other places that you can buy computer hardware besides Best Buy. We have this thing called the Internet, and you can order parts and get them delivered to you.
Realistically, Global Foundries will just keep fabricating chips on the process node that they already have. Plenty of chips don't need the bleeding edge. The problem was that it costs billions of dollars to bring a new process node up, and they simply don't have enough customers to make money off of it.
The bigger question to me is what happens to IBM. IBM paid Global Foundries $1.5 billion basically to take their fabs off their hands and continue development. AMD had already shifted a lot of future development to TSMC. Has IBM done likewise?
Global Foundries had surely told AMD that this was coming well before the public announcement. That certainly played a role in so many AMD chips being developed for TSMC's 7 nm process node. Presumably they've done so with their other customers as well.
AMD used Global Foundries because they were contractually obligated to do so, not because they particularly wanted to. Nvidia could have used TSMC if they wanted to, but didn't.
IBM is more of a services company and systems integrator than anything these days. Sure, you can still buy an IBM branded PowerPC computer - Power9 was available in 2016 on 14nm, and Power10 is slated to be out in 2020 on 10nm. But those are niche systems even among niche users.
IBM doesn't do much with consumer parts, but they do build some mainframes and servers. I don't shop in that market, so I don't know that much of what they do. I do know that they make Power 8 servers with NVlink so that you can attach Nvidia GPUs and get a lot of theoretical bandwidth to connect the CPU to the GPU. Apparently Nvidia tried to get Intel to do that with Xeon parts, and they refused. Of course, given Nvidia's complete inability to get respectable performance out of their PCI Express controllers, I wouldn't be in a rush to convert all of your CPU code from x86 to Power to try it.
Not really that concerning since they were way behind the other fabs to start with.
Of course, rumors of a global RAM shortage could just be conspiracy or price fixing - lord knows the industry hasn't seen any of that happen...
The biggest single reason is not their own volume - in theory there is always contract work (in theory) - its cost. Your link had $10-$15 and a $20+ billion price tags for a new 7nm plants / extensions etc. In other words: a big number. And the one that Samsung broke ground on in 2015 had estimates of $14billion and $15billion. A big number so clearly going to be on these lines. So a big % of what AMD are capitalised at to put it into context.
Edit: I had missed the 2017 story about Samsung expanding the plant they started in 2015 investing another $32billion of which over $5billion is to expand "chip" production (V-Nand). So $20billion for the plant. (Rest is OLED / panels it seems.)
If they manage to do this then they will - almost certainly - agree to use X for any future 7nm production.
Who X might be though - tough question. As I said not TSMC since they have 7nm. IBM pulled out. Not sure either Samsung or Intel would want it either as they have their own investments.
The only one I came up with was SMIC. Big push to get into fabrication, potentially big (government) pockets and also partnered with Imec.
(I thought about Apple as well but - just no. Volume etc. etc plus how long it will take to get up and running. Could be wrong I suppose - who knows with Apple.)
For anyone else::
TSMC are making 7nm it seems using existing litho' techniques - so no reason. Samsung have announced in June that they will be starting 7nm risk production in Q2'18 using EUV - so they are ahead there and Intel ..... are working at it. Would Intel spend serious money just to try and iron out whatever issues they have?
All of which means that a new entrant - if one appeared - will be playing catch up since it will take some time to build a 7nm fabrication plant. For SMIC though I doubt this would matter.
The reason RAM prices are so volatile is that they have to decide years in advance how much capacity they're going to build. It takes years to bring a new fab up, or to refit one designed for an old process node to support a new one. They can't just see that prices are high and ramp up production within a few weeks.
If the industry collectively guesses too low on demand, you get shortages and high prices for a while like we have now. Guess too high collectively and you get a glut and very low prices like the one that drove Elpida out of business in 2012. If price fixing were the culprit, we'd never see the gluts and low prices.
This color change may or may not be temporary.
I expect that Global Foundries' renegotiated wafer sharing agreement with AMD will basically free AMD to go elsewhere for everything beyond 14 nm. At most, AMD will be obligated to continue using Global Foundries for existing chips that already do and maybe some future chips that don't need a state of the art process node, such as chipsets. That could be a good deal for AMD as it eliminates the risk that they'll be forced to use inferior process nodes if Global Foundries struggles--which is exactly what has happened for the last several years.
^Good vid on the history.
Gloflo should remain profitable for a good long while yet, especially as chiplets and interposers come in, but in the long run it seems like the writing is on the wall...
https://www.anandtech.com/show/13277/globalfoundries-stops-all-7nm-development
That they've lost money for so many years in a row is why they decided to stop.
They had three choices. Continue wasting billions on stuff that other companies are already developing better. Merge with one of their competitors and combine resources to better rollout a new process node. Or license the technology from one of their competitors later for a fraction of the billions wasted developing it.