GlobalFoundries: Only a process tech licensee?

[Kuiva maa]

I don't think Zen will be cat class. As a matter of fact I don't see AMD making a new small x86 core at all. Intel big cores are very efficient these days so they have been scaled all the way down to ULV devices and atom is now assigned to tablets etc. Carrizo seems to be following this efficiency trend too, probably Zen will be similar, I am expecting to see some form of it for embedded solutions even.

I don't think AMD will attempt again getting in tablet space, certainly not with x86 so there is no need for a new cat class core. PS4/Xbone successors will probably have similar power envelope and therefore will be served by big x86 cores,be it from intel or AMD. Zen will most likely be a Core competitor.

 

[MiddleOfTheRoad]

You're drawing the conclusions too far. All has said is that they assumes they will get better ROI from other investments, and have to down-prioritize Carrizo desktop due to budget reasons. Anything else is just speculation.

Considering the design objectives of Carrizo (5% CPU performance improvement over Kaveri / with 40% lower power consumption), it really wouldn't have been worth an upgrade as a desktop processor. It makes sense for them to skip it as a Desktop offering.

 

[DrMrLordX]

I don't think AMD will attempt again getting in tablet space, certainly not with x86 so there is no need for a new cat class core. PS4/Xbone successors will probably have similar power envelope and therefore will be served by big x86 cores,be it from intel or AMD. Zen will most likely be a Core competitor.

Don't know about that. Nolan and Amur are both supposed to be tablet-class CPUs. Amur is completely MIA so far (granted, we aren't past any paper release dates ala Seattle), but references to Nolan have shown up in a few obscure locations. The real question is whether or not Nolan will have fundamentally improved Puma+ cores over Beema/Mullins(Puma).

We already know (er well mostly) that AMD has slapped standard Puma cores into Carrizo-L. If Nolan shows up with the same cores then we can conclude that the cat cores are probably dead, or just on the back burner for awhile. Regardless of what we see, it appears that AMD is still making at least a half-arsed effort to enter the tablet space on Intel's coattails with Nolan.

 

[AtenRa]

Changing from a CMT to SMT design means you are investing more logic to the Core for higher IPC (Single Thread Throughput). That means they will have to widen the core (More Integer execution units) for the SMT Thread to have resources to use when the main thread stalls, under-utilized etc).
So even if they are going to use a small core design like the Cat cores, they will have to make them a lot "fatter" (increase the Front-End to allow dual threads, bigger caches, increase the execution units etc).
SMT and CMT in my opinion is not suited for small core designs, it is the reason nobody in the ARM armada have ever used such a design and they preferred to go Big-Little than SMT or CMT etc. Even Intel has stopped using SMT in ATOMs.

 

[Fjodor2001]

Considering the design objectives of Carrizo (5% CPU performance improvement over Kaveri / with 40% lower power consumption), it really wouldn't have been worth an upgrade as a desktop processor. It makes sense for them to skip it as a Desktop offering.

Well that would be on par with Intel for "per generation performance improvements", and better than Intel when it comes to lowering TDP.

 

[Fjodor2001]

This thread has spawned off into quite different directions than originally intended.

Is there anyone that would like to comment on the OP? I.e. can we expect GF to provide their own process tech after 14 nm? I.e. was licensing Samsung 14 nm a one time incident, or will GF be a pure licensee on later nodes too? :hmm:

 

[ThatBuzzkiller]

The way the fab club works is you basically are paying an upfront membership fee in lieu of paying a backside licensing fee, but in terms of how the IP (process node) itself is created, developed, and rolled out it is very much a "licensee" type of business model.

In that regard, Globalfoundries has since its birth been a process tech licensee. IBM's R&D engineers develop the nodes, taking feedback from a handful of GloFo R&D engineers who are stationed in Fishkill but pretty much required to sit on their hands on the sidelines and watch from afar, and then handing off the node to the fab club members so they can internalize it within their fabs and attempt to ramp the yield from near zero to something manufacturing worthy.

The model itself was rather broken, not anything close to the image of collaborative partnership that made it into the PR, but IBM recently had to give GF $1.5B just to convince GF to take their fabs off their hands because even IBM wasn't making enough money off its fab club partners to support the business model. (ergo the abrupt swing towards licensing 14nm from Samsung)

We'll see what happens down the road, I bet Samsung will license 10nm to GF as well. The timing is too soon for GF to absorb the IBM M&A to get their own 10nm stuff off the ground in any better shape than 14XM got off the ground.

But 7nm may see GF roll out their own true internally developed node. If they don't, then it is hard to conceive of what kind of future they will have as a foundry that relies on other foundries for their process tech.

Been Samsung's lapdog is better than having to shut down a business ...

At least on the bright side GF can see a lot more customers than used to ...

 

[krumme]

This thread has spawned off into quite different directions than originally intended.

Is there anyone that would like to comment on the OP? I.e. can we expect GF to provide their own process tech after 14 nm? I.e. was licensing Samsung 14 nm a one time incident, or will GF be a pure licensee on later nodes too? :hmm:

I think idc have given the logical answer to that.
I will just add gf and in reality amd is owned by mubadala. As their motives can change due to more political reasons i think its extremely difficult to assess.

My guess would be that as the original plan to place a fab in the dessert was abandoned there is one less reason to keep gf in the portfolio. But selling it is more than difficult. Not only because its worth nothing but because the wsa clouds roi assessment. Thats a risk. And risk is less money.
On the other hand it seems obvious something must change. So imo either expand or sell. Expansion cost can hide or postpone prior failures so thats an added benefit to it from a political view.
Anyway my guess would be that within the next 4 years we will see some more radical changes going far beyond if they license 7nm or not from Samsung.

 

[Enigmoid]

Well that would be on par with Intel for "per generation performance improvements", and better than Intel when it comes to lowering TDP.

Not really if you consider intel generally gets a power drop across the entire lineup from 1 ghz to 4 ghz rather than only at lower frequencies.