Shootout At 28nm

By Ed Sperling & Mark LaPedus
Samsung, Soitec and STMicroelectronics are joining forces on 28nm FD-SOI, creating a showdown with TSMC and others over the best single-patterned processes and materials and raising questions about how quickly companies need to move to the finFET technology generation.

The multi-source manufacturing collaboration agreement for fully depleted silicon-on-insulator technology means that Samsung will become the second major foundry to offer FD-SOI, following in the footsteps of GlobalFoundries. The Samsung process will be qualified early next year for volume production.

“The 28nm process is projected to last for a very long time,” said Giorgio Cesana, FD-SOI marketing director at STMicroelectronics. “China and other markets are not even at 28nm yet, and this is the last node for single patterning. It’s very appealing for mobile and wearable technology, and it’s the ultimate technology in terms of cost for this level of performance and power.”

For years, Intel and TSMC separately dismissed SOI, saying that bulk CMOS has several advantages over SOI. “We have always done (bulk) CMOS for good reason,” said Jack Sun, vice president of research and development and chief technology officer of TSMC, in a recent interview. “SOI has always been a niche technology.”

In keeping with that argument, TSMC recently expanded its 28nm process lineup with a new bulk CMOS derivative. The technology, dubbed 28HPC, is a low-cost version of the company’s popular 28nm mobile process, called 28HPM. Targeted for mobile phones and IoT applications, 28HPC is positioned at the same price at 28nm PolySiON, but it supposedly delivers 15% better performance. At its recent technology forum, TSMC positioned 28HPC as a lower power solution over 28nm FD-SOI. “To meet customer demand in the mid- to low-end smartphone markets, we have developed 28HPC as a low-cost solution with around a 10% die size reduction,” said J.K Wang, vice president of operations at TSMC.

Added B.J. Woo, vice president of business development at TSMC: “We have tried to drive down the 28HPC Vdd and SRAM Vcc. By taking advantage of better mismatch characteristics of high-k/metal-gate technology, we would like to drive down the SRAM Vcc further around 650 millivolts, which can help drive down the Vdd operating voltage.”

ST’s Cesana countered, saying FD-SOI remains significantly faster and more versatile. “What TSMC offers is limiting the user in terms of what they can do,” he said. “FD-SOI is faster, and it has lower leakage.”

The price of power
The FD-SOI vs. bulk CMOS war is being fought on three fronts. One involves the performance and power characteristics of FD-SOI versus bulk at 28nm. The second involves the performance and power characteristics of both of those versus finFETs at 16/14nm. The third, and historically the most bruising to FD-SOI, involves cost.

FD-SOI has been a more expensive option than bulk in a wafer-to-wafer, which has been an issue for FD-SOI adoption at older nodes where current leakage isn’t a major problem. It would have been significantly more expensive at 16/14nm (using a 20nm BEOL process), as well, had extreme ultraviolet lithography been available. But with commercially viable EUV delayed at least two more process nodes and multi-patterning now a very real option, 28nm FD-SOI has emerged as a lower-cost alternative. The advantages are compelling enough, in fact, that ST says 14 ASICs and 3 ASSPs will tape out in 2014 with volume production to follow soon afterward.

“There is less processing, fewer masks—so you save on the cost of the masks—and the fab equipment is almost fully depreciated already,” said ST’s Cesana. “Add in a larger ecosystem and more volume, and that will lower the cost of the substrate.”

There also are fewer metal layers required in the 28nm process using FD-SOI, which further clouds the price differential.

“With a fully processed wafer, it’s a wash,” said Paul Boudre, chief operating officer of Soitec. “Over the last two years we’ve been working on the process to simplify it. We’ve also increased the supply, which in turns lowers the risk.”

That supply increase was a result of a 10-year licensing deal that Soitec signed with Shin-Etsu Handotai in October 2012, which provides a second source of FD-SOI wafers. Boudre said that was one of the key missing pieces in the supply chain.

FD-SOI is a planar technology that will scale for three generations—28nm, 14nm and 10nm. Following that, the FD-SOI camp has been talking about finFETs on FD-SOI beyond the 10nm node. Christophe Maleville, senior vice president of Soitec’s microelectronics business unit, said FD-SOI also can be extended vertically into 2.5D and 3D designs with either SiGe or III V materials.

Other competitors
GlobalFoundries, TSMC, and Samsung are not the only players in the 28nm foundry game. Though slightly late, SMIC and UMC have separately begun to ramp up their respective 28nm processes. And while Samsung is on board with FD-SOI at that node, it also is offering bulk options at 28nm and finFETs at 14nm.

“FD-SOI is not a replacement of what we offer, it’s an extension,” said Kelvin Low, senior director of marketing for Samsung Foundry. “We expect some customers will still migrate to finFETs. FD-SOI is a compromise on power and performance and cost.”

Low said the addition of FD-SOI at 28nm is similar to the improvements that have been made to older nodes such as 65nm or 130nm, with the price expected to drop significantly as the technology becomes more mainstream. “There are still multiple options available. But what you’re seeing is more companies taking a more collaborative approach. And from our standpoint, we are increasing our position in the foundry business.” In March, Samsung also added 28nm RF capabilities on top of its 28LPP high k/metal gate process, and offers 28LPS (pSiON) and 28LPH versions, as well.

Of the 28nm manufacturers, GlobalFoundries was the first to jump on FD-SOI as a future direction when it agreed to manufacture devices for ST on a foundry basis. The move, in effect, increased ST’s FD-SOI sourcing capacity by complementing its internal manufacturing in Crolles, France with that of GlobalFoundries. At the time, GobalFoundries also positioned itself as an FD-SOI foundry source for customers beyond STMicroelectronics. GlobalFoundries also made SOI-based chips on a foundry basis for IBM, as well.

However, the foundry deal between STMicroelectronics and GlobalFoundries was a non-exclusive foundry deal. For some time, ST was looking for another foundry partner and there was speculation that the company was talking to SMIC, UMC and others. Ultimately, ST and Samsung forged a deal.

Upbeat projections
Virtually everyone sees 28nm as a healthy process node for years to come. In fact, while it isn’t the bleeding edge of design, it’s expected to offer enough benefits of scaling without the added cost of multipatterning or finFETs—which add increased power density, electromigration and ESD threats—particularly when additional power and performance benefits are added into the process or the materials.

“We see very strong demand for 28nm this year,” said TSMC’s Wang. “For 28nm, we tripled our capacity last year. We will continue to increase the capacity by 20% and exceed 1.3 million wafers annually this year.”