Translation from English

Saturday, June 6, 2015

Extreme Tech- Samsung's Plans " A Paper War"?

Is Samsung’s plan to ramp up 10nm production by the end of 2016 a paper war?

Samsung S2 foundry

Share This Article

Samsung has only just begun shipping 14nm hardware in the last few months, but the company is already aggressively talking up its plans for a 10nm ramp in 2016. In fact, this is the second such push we’ve seen in recent weeks — TSMC is talking up its own 10nm plans as well. Ironically, the one company being quiet about its own 10nm ramp is Intel, which has said only that it will detail its own push later this year. Of the three, Intel has the strongest track record of delivering foundry nodes before the rest of the industry, and tends to lead with its adoption of next-generation technologies. Samsung, however, may be gearing up to spend as much as $8 billion on the cutting-edge node.
The war of words between Samsung and TSMC, however, is just that so far — a war of words. We’re already seeing rampant debate and speculation over whether Qualcomm or Apple will build with either foundry at 10nm, despite the fact that neither firm has shipped a single chip on 14nm yet. The new argument we’ve already seen floated is that 14nm, like 20nm, may be a new “lite” node, with full standardization at 10nm.
Over at SemiWiki, contributor Rmaire argues that this rapid-fire cycle is being driven by smartphone vendors who have tied the evolution of hardware and software platforms to a yearly cadence that pushes the foundry business to standardize on new nodes as quickly as possible, rather than adopting a longer business cycle. Apple’s experience is instructive in this case: The Apple A4 debuted in 2010 at the 45nm node. The Apple A5 launched on 45nm in 2011 but moved to 32nm in 2012. The Apple A5, A5X, and A6 were all produced on 32nm through 2012, but the Apple A7 moved to 28nm in 2013 and the A8 was built on 20nm in 2014. This year, the upcoming Apple A9 is widely expected to launch at 14nm.
Apple isn’t the only company that tends to move to cutting-edge processes and nodes, but it and Qualcomm are easily the largest design firms with such an interest in high-end manufacturing. Unfortunately, this relentless cadence has consequences. One, that we’ve already seen, is that TMSC and Samsung are engaged in a relentless game of PR one-upmanship. The other is that these new nodes may become “lite” precisely because the companies that can invest in the latest and greatest production are being shoved towards ever-smaller nodes, which only exacerbates the problem of amortizing and making money in a given product node over time.

Upending the traditional model

In the traditional foundry model, manufacturers like TSMC, Samsung, UMC, and SMIC make a significant percentage of their total profits on nodes that companies like Intel abandoned ten years ago. This graph, from TSMC’s Q1 2015 presentation, shows how 39% of the company’s revenue is driven by process nodes that are 9-20+ years old. If we include 40/45nm (45nm debuted in 2008), then TSMC earns just over half its revenue on technology nodes that Intel, AMD, and Nvidia haven’t used in years.
TSMCRevenue
Foundries typically pay for new nodes by attracting new customers as costs drop and performance improves. The problem is, that’s not happening at newer nodes the way it used to. In some respects, it’s not going to happen at new nodes, period. If you have to use double patterning to work with a new node, you’re going to wait much longer before adopting it — assuming you adopt it at all.
We’re already seeing an increasing number of companies sticking with larger nodes — UMC and ARM recently signed a deal to advance 55nm technology for IOT because that node offers a variety of cost and power consumption advantages. While higher-end devices are implemented on lower nodes, a modest IOT device can achieve a leakage power reduction of up to 350x (not a typo) by using new approaches to these “classic” node designs.
ISS_F1
ISS_F1
If you’re a modern foundry, however, these trends are problematic and could feed a bipolar approach to the problem. On the one hand, foundries like TSMC and Samsung have every reason to bang the drum regarding new nodes, pushing customers and device designers to think about their upcoming 14nm and 10nm projects and to compete for design wins from the handful of giants. At the same time, there’s good reason to continue refining older nodes — if you can squeeze an additional 15-25% performance out of a node between 55nm – 28nm, you may be able to convince firms to bring up new chips on it and win higher margins for the legacy node, offsetting some of the costs of the new technology.

An uncertain impact on graphics and a note on foundry ramps

One of the major unknowns, at this juncture, is how these trends will impact the graphics market. High-powered SoC designers were the big losers at 20nm — no one built process nodes that were suitable for next-generation GPU implementations, and as a result, both AMD and Nvidia skipped them. Both companies are standardizing on 14/16nm, but AMD and Nvidia combined don’t do a tenth of the sales volume of an Apple or a Qualcomm. While GPUs have historically been important to process ramps, that importance may be fading, replaced by yearly cadence cycles from smartphone vendors who can command sales that no graphics card can hope to match.
Finally: Despite big promises from TSMC and Samsung, Intel and its foundry competitors mean very different things when they talk about volume production on a given node. When Intel says it’s begun volume production on, say, 14nm, it means that chips using that technology will be in shipping hardware within a quarter, two at the outside. Intel can deliver on those kind of timetables because it performs so much of the validation and qualification work in-house, and because it times its own milestone announcements to correspond with its OEM partner’s shipping dates.
Companies like Qualcomm and Samsung, in contrast, operate on very different timetables. When Samsung ramps a process node, it typically enters volume production on DRAM or NAND first, followed by logic at a later date. Getting a Qualcomm or Apple SoC ready for volume production is just the first step. Once validated, chips still have to be shipped to various network operators for validation on their own networks. It can take anywhere from 5-9 months for carriers to perform their own validation or device customization work. This is why Samsung can say it’ll be in 10nm volume production by the end of 2016, yet we may not see devices based on 10nm hardware until Q2 2017. GPUs do tend to be a bit faster — typically AMD and Nvidia have launched hardware 1-3 months after TSMC announces volume production at a given node.
As we’ve seen at 14nm, manufacturing troubles and R&D decisions can sway which company leads at a given node. Intel still holds an overall lead on feature size and first node deployment, but Samsung beat TSMC as the first pure-play foundry to hit that feature size. Which company will lead at 10nm and how relative densities and cost structures will compare is still very much unknown — most vendors are still finalizing 14nm designs, not rushing to leap for a new node.
Comment

Post a Comment

Posted by at

No comments:

Post a Comment

Please leave a comment-- or suggestions, particularly of topics and places you'd like to see covered

Subscribe to: Post Comments (Atom)