AMD server roadmap, 28nm in H2 2014

[mrmt]

Don't trust AMD Roadmaps so often particularly the Enterprise/Sever roadmaps.

I agree with you that it is not a good idea to trust AMD roadmaps, but I don't think this leak should be completely discarded. Each roadmap revision AMD did brought bad news, we got either delays, or cancelled and watered down designs, never a revision for good. So this roadmap indicates the best case scenario for AMD.

I'm eager to see what AMD is going to field in 2014. When Intel launched SNB-EP Intel toasted AMD server business because the chip was much superior to Bulldozer and 32nm was already very mature and winding down on consumer so they could fill all the relevant price points. We are going to see the same in 2014 but in a much worse fashion. Intel launching Haswell-EP on a very mature 22nm process against whatever AMD throws at it on 28nm.

 

[ShintaiDK]

It's really amazing. I know that GF has already run various test packages @ 20nm, and read a rumor somewhere of them running some ARM SoCs (no word on how they came out) and yet we'll not see a high performance 20nm product till 2016

You mean the "less than 12000 transistor" Cortex M0 testchip? And you wonder why there is no products?

 

[Ajay]

This isn't even going to fun to watch. It'll be sad, very sad. NostraSeronx's road map had better be what AMD does. Heck, even if they only went with a die shrunk SR @ 20nm, at least AMD wouldn't be totally out of the game on Perf/Watt, just badly behind. At 28nm, forget it.

 

[Ajay]

You mean the "less than 12000 transistor" Cortex M0 testchip? And you wonder why there is no products?

No, I don't wonder now
Thanks.

 

[NostaSeronx]

The road map depends on the WSA2013 agreement with GlobalFoundries. If AMD can get fixed wafer pricing and for a good price you can bet your behind you will see Dublin, Kaveri, and Macau in 2013.

 

[Ajay]

The road map depends on the WSA2013 agreement with GlobalFoundries. If AMD can get fixed wafer pricing and for a good price you can bet your behind you will see Dublin, Kaveri, and Macau in 2013.

What are Macau and Dublin? BD derivatives I'd guess, but exactly what type of derivatives? BDv3 or something else?

Hmm, GF is still building up capability @ it's NY Fab : http://www.indeed.com/q-Globalfoundries-l-New-York-jobs.html

 

[NostaSeronx]

What are Macau and Dublin? BD derivatives I'd guess, but exactly what type of derivatives? BDv3 or something else?

http://www.xbitlabs.com/images/news/2011-08/amd_server_roadmap_2011_2013_august_2011.jpg

If:
Terramar, Sepang, Komodo, Trinity = Piledriver

Then:
Dublin, Macau, Komodo(Unknown), Kaveri = Steamroller

 

[Arachnotronic]

32nm SNB already mops the floor with 32nm "Piledriver". 22nm IVB server part will only turn the screws tighter, and then finally a Haswell-EP part with some seriously cool uncore goodness will put an end to AMD's "server" attempts.

AMD needs to focus on something. Servers ain't gonna be it.

 

[Ajay]

http://www.xbitlabs.com/images/news/2011-08/amd_server_roadmap_2011_2013_august_2011.jpg

If:
Terramar, Sepang, Komodo, Trinity = Piledriver

Then:
Dublin, Macau, Komodo(Unknown), Kaveri = Steamroller

But that Roadmap is over a year old!

 

[pablo87]

With AMD, the (road)map is not the territory.

 

[Idontcare]

What are Macau and Dublin? BD derivatives I'd guess, but exactly what type of derivatives? BDv3 or something else?

Hmm, GF is still building up capability @ it's NY Fab : http://www.indeed.com/q-Globalfoundries-l-New-York-jobs.html

Strong rumors are pointing to GloFo pulling out of the IBM consortia, taking R&D development internal, and moving the R&D work to it's NY fab. Apparently GloFo is already in the process of relocating key R&D personnel to internalize the program but they haven't announced anything to the masses yet (neither internally nor externally). Could be an interesting 2013/2014 for GloFo.

 

[Olikan]

Strong rumors are pointing to GloFo pulling out of the IBM consortia, taking R&D development internal, and moving the R&D work to it's NY fab. Apparently GloFo is already in the process of relocating key R&D personnel to internalize the program but they haven't announced anything to the masses yet (neither internally nor externally). Could be an interesting 2013/2014 for GloFo.

why this would this be a good thing?

 

[Ajay]

Strong rumors are pointing to GloFo pulling out of the IBM consortia, taking R&D development internal, and moving the R&D work to it's NY fab. Apparently GloFo is already in the process of relocating key R&D personnel to internalize the program but they haven't announced anything to the masses yet (neither internally nor externally). Could be an interesting 2013/2014 for GloFo.

Interesting.

I thought that the amount of talent they were looking for was a bit excessive. This does fit with GloFlo's plans to be somewhere in between an IDM and a Foundry - as that's the only way they see Foundries surviving (adding value for their customers). I watched a video where they were talking about offering standard IP blocks (circuits) optimized for various parameters (performance, size, low power, etc.). If I had a freaking memory, I'd link it.

 

[Idontcare]

why this would this be a good thing?

Working in the fab club is beyond frustrating for the actual boots on the ground there. There is only one way things are allowed to be done and that is the IBM way. Member affiliates don't get equal say in decisions, it is not a partnership, it is a hiearchy where partners get told what the decisions are and then the partners have to spend time fixing the broken process once its transferred to their own fab. GloFo's 32nm lost about 6 months due to a BEOL dielectric cracking issue that the GloFo R&D engineers at Fishkill highlighted was going to be a problem nearly a year in advance, but they weren't taken seriously by their IBM counterparts until it was too late and the broken process transferred to Dresden (where the Dresden engineers had to then fix it insitu).

 

[podspi]

IDC, is this just because IBM doesn't care, or because IBM's goals for the process are different from most other fabs?

Just seems unlikely to me that IBM would just develop a 'broken process', but maybe I'm wrong. I know very little about IBM's fab business.

 

[Idontcare]

IDC, is this just because IBM doesn't care, or because IBM's goals for the process are different from most other fabs?

Just seems unlikely to me that IBM would just develop a 'broken process', but maybe I'm wrong. I know very little about IBM's fab business.

IBM has a history of this, this is not the first instance, nor will it be its last.

Altera pounces as Xilinx becomes latest to abandon low-k SiLK: Rivals scrap in dielectric flap

^ that was 2003, when literally every other IDM in the world knew SiLK was broken and could not be made manufacturable. And yet IBM arrogantly pursued it anyways, to the detriment of their (then) fabless customers.

The reason this happens is because there is a big disconnect internally at IBM between the decision makers for critical process node decisions versus the individuals that are tasked with waving magical wands and somehow making it all work. The cracking dielectric uncovered in packaging was known to IBM, but they were OK with their 32nm timeline slipping until it was fixed. (what's a year among friends?)

It was GloFo and AMD that got hosed by IBM's malaise. What made it even more painful for GloFo was that it was they who raised the flag up the pole as early as possible to highlight the issue and yet they were more or less told to sit on it and rotate.

See many 45nm HKMG processors from IBM's fabclub consortia members?

 

[mrmt]

IDC, is this just because IBM doesn't care, or because IBM's goals for the process are different from most other fabs?

The latter probably. You just have to look at what kind of chip IBM makes to see that it makes a lot of sense for them to use expensive techniques such as SOI.

For AMD, as it didn't have neither the knowledge nor the money to develop a process in-house get an inadequate IBM process should have been worth for a lower R&D bill. Now that GLF is looking far beyond AMD, it doesn't make sense to use IBM process as they must have something tailored for their customers, otherwise they will never get a top tier client. The only other bleeding edge foundry where they could get this knowledge, TSMC, won't share their R&D with them, so they have to pull out and develop their own process.

 

[Ajay]

Wow, some great comments there IDC & mrmt.

I remembered the delay, but not why and definitely didn't know the problem was identified a year in advance! Geez, AMD couldn't catch a break (pun intended).

Looks like GloFo has learned it's lesson. The state of NY must be ecstatic over the additional high-end engineering hires! They are not only going to develop their own process tech, but are also going to design they're own circuit blocks to make available to their customers (pseudo-IDM). That's going to be quite a Fab and Design center in Malta, but they have allot of work to do to catch up with TSMC.

I wonder what Samsung plans to do (especially if they lose cash cow Apple to Intel as per current rumors). Although they have allot of internal demand to fund their IDM efforts, I don't know if they are still depending on IBM's process tech.

 

[Olikan]

nice info IDC makes everything more clear
i was thinking that IBM there could only help, since the have a TONS of IP

 

[Idontcare]

I wonder what Samsung plans to do (especially if they lose cash cow Apple to Intel as per current rumors). Although they have allot of internal demand to fund their IDM efforts, I don't know if they are still depending on IBM's process tech.

Samsung is in the fab consortia solely to keep an eye on what is going to be made available at the other foundries (excepting for TSMC of course). Samsung actually doesn't use the IBM fab club process tech, they have their own internal R&D pipeline in Korea that develops the process node they end up putting into their production fabs.

In a sense, the IBM fab club enables legal corporate espionage provided you are willing to pay the membership price. Being a member of the Sematech consortia is no different (we, TI, were a member too, for pretty much the same reasons Samsung is a member of the IBM consortia).

 

[Abwx]

Here's the statement we received this afternoon from AMD's Chris Hook:
AMD has a long history of supporting the DIY and enthusiast desktop market with socketed CPUs & APUs that are compatible with a wide range of motherboard products from our partners. That will continue through 2013 and 2014 with the "Kaveri" APU and FX CPU lines.

http://techreport.com/news/24009/amd-were-not-abandoning-socketed-cpus

 

[Ajay]

Samsung is in the fab consortia solely to keep an eye on what is going to be made available at the other foundries (excepting for TSMC of course). Samsung actually doesn't use the IBM fab club process tech, they have their own internal R&D pipeline in Korea that develops the process node they end up putting into their production fabs.

In a sense, the IBM fab club enables legal corporate espionage provided you are willing to pay the membership price. Being a member of the Sematech consortia is no different (we, TI, were a member too, for pretty much the same reasons Samsung is a member of the IBM consortia).

Sounds like the Interop events I used to go to in Networking, yeah, we'd fine tune our code to play well with others, but we also got to see peoples R&D pipeline and even code at times. Like you say, legal corporate espionage.

As far as Samsung goes, looks like they are already in a better place than GF. I guess we'll see how deep ATIC's pockets are in the race to be a top foundry partner.

 

[Idontcare]

As far as Samsung goes, looks like they are already in a better place than GF. I guess we'll see how deep ATIC's pockets are in the race to be a top foundry partner.

Samsung has issues of a different sort. Their legal team out-ranks their foundry division, so when samsung legal says "let's sue our foundry customers, even though it means it will kill all future foundry work" that becomes the only decision that matters and the foundry guys (samsung employees working in the foundry division) get hosed despite their desire to be a good foundry partner.

Right now no fabless IC design house wants to put their designs into a samsung fab because Samsung has done nothing to distance itself from the perception that they will just steal your IP to build their own clones of your products and then compete against you in short order, and if you mount any legal resistance then they will bury you with a legal dept that dwarfs your entire R&D expenditure as a customer.

The samsung foundry guys are not happy that they've basically had their future business turned against them because their mobile division (completely managed by an entirely separate team) convinced the legal team that there was more money to be made in just competing against Apple rather than making money being their foundry.

No fabless company wants to see themselves become the next victim of that sort of predatory "bait-and-clone" foundry business model.

 

[beginner99]

The samsung foundry guys are not happy that they've basically had their future business turned against them because their mobile division (completely managed by an entirely separate team) convinced the legal team that there was more money to be made in just competing against Apple rather than making money being their foundry.

No fabless company wants to see themselves become the next victim of that sort of predatory "bait-and-clone" foundry business model.

White I hate lawyers and this business practice is morally not very acceptable it's probably true they make a lot more money selling complete phones than just SOCs and even if you factor int he costs for the following battles with Apple.

 

[mrmt]

No fabless company wants to see themselves become the next victim of that sort of predatory "bait-and-clone" foundry business model.

The clone part isn't exclusive to semicon. With Samsung everything from cars, appliances, TVs and everything else rely on some kind of cloning at some part of the product history in the group.