Last week, Samsung and GlobalFoundries announced that they would partner together to deploy Samsung's 14nm technology at all GF facilities. The deal is unprecedented in modern foundry history, with GF essentially acknowledging the two companies will use a "copy-smart" approach that involves synchronising materials, process recipes, and tools.
That's not quite the level of duplication that Intel's "copy exactly" approach uses, but it should still allow designs built at one foundry to ramp up smoothly at another.
The elephant in the room, however, is the company that wasn't mentioned – IBM. Ever since GlobalFoundries was spun off from AMD in 2009, Samsung, IBM, and GF have maintained what they called the "Common Platform." Common Platform was an agreement on broad technology standards that would be deployed across all three companies.
It was never entirely clear which technologies were governed by CP rules and which were independently developed, but the goal was to create an infrastructure that allowed for easy design porting across all three corporations.
Now, Samsung and GF are getting married at the 14nm node and IBM is nowhere to be seen. We've previously reported that the company is exploring a complete exit from the semiconductor business; news since then has suggested that GlobalFoundries is actually the leading contender to acquire IBM's manufacturing capabilities. What that means for long-term R&D is unclear – while IBM manufactured very low chip volumes, it has continued to do a great deal of leading-edge R&D.
GlobalFoundries 14nm-XM is mostly dead
GlobalFoundries hasn't officially declared that its hybrid 14nm-XM process is pushing up daisies, but all signs point in that direction. With 14nm-XM, GlobalFoundries claimed a 20 per cent performance boost at constant power or a 40 per cent power reduction at the same performance. Samsung's 14nm Low Power Enhanced (LPE) node claims a 20 per cent performance gain or a 35 per cent power reduction – a shift of just five per cent.
We've heard persistent rumours that GF's 20nm and 14nm-XM had run into delays and difficulties, and this strategic realignment with Samsung would seem to support that conclusion. Customers reportedly weren't sure if GF's approach, which combined a 20nm back-end with 14nm front-end processing, would give them the benefits they were looking for. Power and performance would have risen under the hybrid model, but die size would've been essentially flat between 20nm and 14nm-XM.
GF's troubles with 14nm deployment, however, probably aren't the only reason for the move – copying Samsung's technology gives GF the option to second source for the Korean company, which could help GF keep its own fabs running at high utilisation. It's also an example of just how difficult it is to ramp up modern foundry nodes – we're seeing multi-billion dollar firms collaborating at this point, because the difficulty just keeps rising.
How long till market?
One of the most frustrating things about covering foundries in journalism is that they tend to fudge their process node launch dates. When Intel, AMD, or Nvidia says "We're now shipping Node X for revenue," it means that new chips built on a new node are one to three months away. Foundries, in contrast, will claim to have begun "volume production" a year or more before actual shipping hardware is available on the market. The type of hardware matters a great deal as well.
At 28nm, TSMC announced volume production in late October 2011 and AMD shipped the first GCN video cards in January of 2012 – a thoroughly reasonable three month gap. The first 28nm Snapdragon S4 processors didn't start shipping in phones until early May – seven months after TSMC began its own volume shipments. Even then, supplies were constrained and Qualcomm publicly complained about low product volumes.
Part of this isn't TSMC's fault – SoCs go through extensive validation and must be tested first by Qualcomm and then by carriers to ensure that products are ready for primetime. Part of the problem, however, is related to how foundry roadmaps are communicated.
The above chart from IBS shows just how long it takes to move a chip from library and IP qualification through to validated prototypes. The length of time is given in weeks, and shows design times moving from 109 weeks at 28nm (a bit over two years) to more than three years by the 14nm node. Bear in mind that this chart assumes that the process node is finalised and ready to go – companies can start ramping their architectures before the process node is fully implemented, but they're building to hit a moving target.
The long lead times explain why TSMC, Samsung, and GlobalFoundries are all talking about 14/16nm ramps in late 2014 to early 2015, but Qualcomm doesn't expect to ship 20nm SoCs until the first quarter of 2015. Broad industry consensus is that 20nm will be a short node with a limited lifespan, but that's only going to be true for companies that are already bringing 14nm designs along now.
Will this hurt AMD?
Probably not. In its last conference call, AMD's Lisu Su, VP and General Manager of Global Business Units, told David Wong of Wells Fargo that, "we are 28 this year, we have 20-nanometer in design, and then FinFET thereafter." Looking at that timeline above, it's obvious that AMD needs to already be thinking about its 14nm designs – especially since these are the first chips that Jim Keller will have had a direct hand in building – but the architecture probably isn't at a stage where GF's roadmap shift will harm AMD's ability to execute.
With GF and Samsung now joining forces at 14nm and Intel reportedly pushing its own 14nm production into the tail end of 2014, TSMC could wind up a winner if its own 14nm production continues on schedule. Nothing short of all Intel's fabs burning down simultaneously would close the two year gap between the two companies, but the Taiwanese manufacturer might shrink the delay down to 18 months.