Tuesday, December 30th 2014
AMD to Switch to GlobalFoundries' 28 nm SHP Node in 2015
Faced with continuous development roadblocks with TSMC, AMD is reportedly planning to switch to the 28 nm SHP process of GlobalFoundries, to build GPUs in 2015. The 28 nm SHP (super high-performance) node will allow the company to lower voltages, giving it greater room to increase clock speeds of its upcoming GPUs. AMD's GPUs in 2015 could be based on its latest Graphics CoreNext 1.2 architecture, and AMD needs every means to minimize voltages, and crank up clock speeds.
The company hasn't abandoned TSMC completely just yet, with reports speaking of AMD using the Taiwanese fab's 16 nm FinFet node to manufacture its next-generation "Zen" CPUs. Zen is the successor to AMD's "Bulldozer" architecture and its derivatives ("Piledriver" and "Steamroller.") It could feature a radically different core design.
Source:
BitsandChips.it
The company hasn't abandoned TSMC completely just yet, with reports speaking of AMD using the Taiwanese fab's 16 nm FinFet node to manufacture its next-generation "Zen" CPUs. Zen is the successor to AMD's "Bulldozer" architecture and its derivatives ("Piledriver" and "Steamroller.") It could feature a radically different core design.
48 Comments on AMD to Switch to GlobalFoundries' 28 nm SHP Node in 2015
For the server sector they will have both Zen and K12 which is 64-bit ARM CPUs that will cover many aspects of the commercial segment.
How can you be so certain the performance of Zen will be so huge that it will close the gap with Intel on both the Enthusiast and Server markets. I know you didn't say it directly but you hint at it. Everyone thought Bulldozer was going to be as revolutionary as the AMD64 was when it first came out, "It will be worth the wait". I wouldn't hold my breath. AMD has a history of delaying products and over promising while under delivering. Not saying it can't happen but I find it to be not as huge as some make it out to be.
Smaller node gives the options for better product segmentation - which is a fairly important factor in securing markets, as evidenced by Intel being able to dial in SKUs tailored for individual clients ( the E5-2666v3 for Amazon and the E5-2698Av3 for Lenovo for example). The smaller node also allows a company to dial in a wider scope of area-ruled design, power/perf-ruled, or the option of packing in a higher transistor count - subsequently Intel's 22nm fabs currently churn out sub-70mm SoC's, 662mm^2 Haswell-EP's and just about every conceivable SKU in between. So, AMD are ahead by a year, but are two years away from a new architecture, while Intel has Broadwell waiting in the wings and Skylake to arrive before AMD's parts....and that's assuming that AMD's product timetable is on time - hardly a given based on AMD's past/current track record. You neither know either of these points as facts, nor are privy to the processor landscape (performance, pricing, availability) that these parts will have to compete in - unless you somehow know the performance of the Intel SKUs that Zen will be expected to compete against.
Your view is exactly the same as the rose-tinted outlook that AMD touted for Bulldozer in 2007 and 2008 when they expected (for some outlandish reason) that Intel would continue to field Core 2 Quad and that was the architecture Bulldozer would compete against. Good luck with that. The micro-server market seems more of a smoke and mirrors exercise than anything concrete, and since Intel own ~98% of the x86 server market, AMD really need not just a compelling product, but a comprehensive platform with enough upside to make server end users trash their Intel-based systems. Legacy products.
From Mercury Research (pdf) latest publicly accessible server figures:
• With a 97.8% unit share Intel continues to dominate the server processor market. AMD has publically stated its interest in developing ARM-based server processors. We think that AMD might at some point choose to discontinue development of future x86 server-specific processors.
The preliminary release of the Mercury Research data did not include average selling price (ASP) information. For the March quarter, Mercury Research data implies that Intel’s server processor ASP (excl. Itanium) was $625 (versus $599 in the December quarter) and that AMD’s server ASP was $326 (versus $353 in the December quarter). • Intel said that its Data Center platform ASP increased 3% sequentially and increased 11% year over year in the June quarter.
Unless they either address the cache latency or somehow make their branch prediction magical they will still be slow, Jorge has no idea other than he seems to have a hard on for spouting how great AMD is, and copy pastas the same crap everywhere, so either he is a paid for AMD shill, or has a severe complex that requires a large amount of medication that is not taken.
AMD is a long way from screwed however, they have already made HSA work, and are offering a one chip solution that neither other player can offer, plus their continued advance in the low end market by choosing to make less has a lot going for them, they just need to get more PR/Advertisement about that segment.
Or............Just another flame war in general.................
Ah its funny as by just seeing someone's forum name you can already know what's going to be typed.:slap:
On point I hope amd do very well with there decision and climb on top again where they belong.
. . .
Realistically looking AMD's original architectures have been worse off performance/efficiency wise compared to Intel more often than not.(the only notable exception being the Athlon series) But that has always been OK because AMD managed to compete incredibly well based on their awesome pricing. I dunno about you, but back in the days the only Intel CPUs I could get my hands on that were in my price range were crabby cut down Celerons...
Nowadays however Intel's manufacturing technology has evolved so far that it has become incredibly cheap for them to release decent CPUs while still maintaining their expected profit margins and AMD are finding it extremely hard to compete using their old tactics.
On a side note if a process is called "14nm", it doesn't mean that every part of the transistor is 14nm, it only means that some parts of a component can scale down to 14nm. I wouldn't be surprised if Intel's 14nm process and GF's 28nm have particular components that are quite similar in size.
Intel Broadwell-Y (14nm FF) : 1.3 billion transistors in a 82mm² die area = 15.85 million per mm²
AMD Kaveri (28nm SHP) : 2.41 billion transistors in a 245mm² die area = 9.837 million per mm²
And well, I've been saying for quite a few years now that AMD needed to make this move. So it pleases me that AMD finally "listens".
That's a different server sector you're talking about...ARM cannot compete with 18-core Haswell CPU's or even with the 12 core versions.
And the reason why intel isn't making huge improvements is because they don't have have, they are competing with their own CPU's. If they wanted they could just push out a 130W CPU on mainstream platform and be done with it.
AFAIK Intel has 10/7nm on roadmap and IBM has some prototypes on sub 10nm process.
The problem I keep seeing is that no manufacturer actually uses APU's in their laptops or should I say everyone prefers Intel (even if that means a more unbalanced setup).
For anyone interested in process tech, here's a link for Chipworks previewof the recently concluded 2014 IEDM. Well worth a read, and follow up.
That makes AMD's "slower on process tech" a positive, not a negative.
I think you'll find that my posting in this thread has consisted of:
Providing linksthat Intel was indeed selling 14nm chips, as a poster had specifically asked for proof.
Disagreeing with a poster that AMD's next big chip will be on a 20nm nodesince just about every piece of evidence points to the processes not being suitable for large, fast, and power hungry IC's, and alerting someone to the fact that you need not estimate performance of GloFo's 28nm SHP process because it is being used in a mature shipping range of APUs
Disagreeing with Jorge over the relative merits of two architectures that haven't seen the light of day, but which Jorge supposedly has enough information to form a solid analysis...and debunked some FUD about how well AMD's server division is doing ( Just because there are Opteron systems in use it doesn't translate to continued meaningful sales, just as being able to buy 429 rebuild kits and spares means that FoMoCo still sells Torino's and Ranchero's)
A back-of-the-napkin comparison of Intel's 14nmFF and GloFo's 28nm SHP process.
I don't think I referenced AMD's "slower on process tech" in any of the posts and doesn't really affect them aside from financial aspects of their WSA's- if anything, posters turned this into an AMD vs Intel flamefest from the off, when the real story should have been Globalfoundries vs Intel vs Samsung vs the pure play foundries. I'm sorry if attempting to offset some OTT guerrilla marketing with some factual counterbalance is somehow worthy of censure.:)
FWIW, I also wouldn't put 10 years of silicon development into the "who cares" category. The past ten years have seen us move from 90nm Prescott P4's and Athlon 64's in addition to 130nm X800's and 110nm GeForce 6800's. Ten years represents a lot of history and a lot of future potential in the semicon industry. There is also an excellent chance that the business landscape and the players involved will not remain as it is today. How do you arrive at that doomsday scenario? GloFo's 28nm SHP was not only tailored to AMD's requirements, but has been successfully applied to GPU related products for the past year so I doubt it represents any kind of gamble. Also not too sure how this is supposed to crater Nvidia's share price - even if Fiji and Bermuda are the second coming of Christ. How would it be any different to 2009-10 when AMD fielded an all new landmark Evergreen series and Nvidia was still dicking around with a two-generation-old G92?
A new range of 28nm GPUs isn't suddenly going to overturn Nvidia's 80% workstation market share, its 85% HPC GPGPU share, or its mobile GPU share (for which AMD's large die GPU seems somewhat unsuited for in any case) - and with $3.2bn in cash on hand and an ongoing share buy-back scheme in place it seems unlikely that the share price will crater because of some desktop GPU business from AMD*. Even if Nvidia had nothing for 2015 - no GM 200, no GM 206, they'd still spin Pascal and Volta, and they'd still have the pro market ecosystems - no different to the situation in September 2009, except now they have high profile OpenPOWER contracts and some SoC automotive stuff going on.
If 28nm SHP were the beginning of some long term advantage I could understand the viewpoint, but we both know that the node is a stopgap until TSMC's 16nmFF+ and Samsung/GloFo's 14nm-XM arrive. So, unless you see 14nm being demonstrably better than 16nm and Nvidia being completely locked into TSMC rather than shopping around to Samsung, GloFo, or UMC, it seems that the playing field will again be in flux in late 2015.
* Quite possible that Nvidia's share price drops to 2004-05 levels ($5-7), but IMO that would require a something with the ability to hit harder than AMD - say, the lawsuit/countersuit business with Samsung/Qualcomm and loss of emerging SoC markets.
I'm not sure how so many people overlook this. Perhaps it is because they have never been a sysadmin or similar supportive role. Great point. I am still rocking my 7970. Perhaps next year I'll think about upgrading.
We will reach the smallest node we can reach before we have the next substance ready to step in, and this is all due to quantum tunneling effects and how many electron volts it will take to activate a gate VS how many can breach the insulation pattern to run to the drain. Its then that better design practice will start making a difference, and when we start learning to use different doping elements to lower the capacitive threshold that is the biggest weakness in current transistor design, and from there I am guessing it will be down to the distance electrons have to travel VS the speed of light in the substance that will be the next threshold, hell we are already having to plan for it.
Also they are already talking about atomic level transistors for the next 10 years, so no worries. :)
Working prototypes are already tested since 2012
(
The benefit comes from a design with the size in mind, maximizing the efficiency of the die to use the tooling available, look at what Nvidia did with the 9xx series VS the shit AMD has with the 295, the 295 is a turd while the 9xx series is simply amazing. Same process node, same basic technology, but Nvidia simply designed a much, much more efficient core in every metric.