Qualcomm Shies Away From High-k At 28nm

Strain engineering can offer similar benefits for less cost; at 22nm HKMG will be a requirement.


By David Lammers
Qualcomm CDMA Technologies said it will not use a high-k/metal gate (HKMG) process for most of the chips it makes at the 28 nm node, sticking with a poly/SiON gate stack. The company described the rationale behind the strategy, which because of Qualcomm’s size will have a major impact on the foundry business, at the 2010 International Electron Devices Meeting (IEDM) held in San Francisco.

Jim Clifford, senior vice president of operations, said the decision to stick with a poly/SiON process was related to timing and cost issues. “High-k inherently requires more mask layers, and there are defect density issues that make it a little more challenging,” Clifford said in an interview at IEDM.

Qualcomm is not closing the door on HKMG. “There is a class of products where you need it,” Clifford said, including some chips made for tablet computers and the “extremely high-end” smart phones. Qualcomm may use the HKMG processes at the later stages of the 28 nm node for those products, which require roughly 2GHz and higher frequencies. But for most of its high-volume smart phone chipsets, Qualcomm will stick with the less expensive poly/SiON process.

The company isn’t alone in taking that route. Texas Instruments is following a similar strategy, according to industry sources.

In a luncheon keynote speech at IEDM, Clifford emphasized that while Qualcomm is hungry for smaller and faster transistors, he is growing increasingly concerned about the costs involved with EUV lithography and other technologies required to stay on a Moore’s Law cadence. “I’m all about costs,” Clifford said.

In an IEDM paper on its 28nm technology, Qualcomm technology director P.R. Chidambaram said a high-k process without significant strain on the channel does not offer an advantage over a poly-SiON process with strain techniques. “A HKMG process with strong strain engineering offers a substantial speed gain but with a higher cost, making it more suitable for smartpad/tablets and extreme high-end smartphones. A poly/SiON process enables a quicker time to market with less process risk, and historical defect density reduction,” he said.

Most of the Qualcomm smart phone chipsets, based on the company’s Snapdragon processor core, run at 1GHz or less, and can be served by going to dual-core designs. Geoff Yeap, senior director of technology, said the Snapdragon-based chips sold by Qualcomm “have huge volumes,” adding that the high-k processes at the major foundries “are not ready yet.”

Yeap said Qualcomm will move a limited number of products to the HKMG foundry processes at a later point. While HKMG transistors generate higher drive currents from the increased inversion charge, the switching capacitance increases due to the need to charge up the gate to gain the higher drive current. Linear drive current (Idlin) is more important than saturation current (Idsat) for Qualcomm, he added.

Although HKMG helps out with power leakage at the gate, it doesn’t do much to improve power lost at the substrate and in the source-drain areas. Qualcomm has been able to contain power consumption by using well biasing, in addition to circuit techniques such as power and clock gating. A specially designed power-gating switch was described by Chidambaram, who cited it as an example of the close cooperation between Qualcomm technologists and engineers at an unnamed foundry partner.

The 28nm node offers major advantages to Qualcomm, which combines RF, digital processing, and analog functionality on its SoCs. Moving from 45nm to 28nm design rules supports a 2.4 times increase in gate density, a 55% reduction in power and a 30% gain in frequency, said Chidambaram.

Qualcomm calculated the gate leakage percentage contribution to the total leakage in various operation modes. Source: 2010 IEDM paper.

Qualcomm calculated the gate leakage percentage contribution to the total leakage in various operation modes. Source: 2010 IEDM paper.

At IEDM, some technologists expressed surprise at Qualcomm’s decision to stick with a largely SiON process, citing better electrical control of the channel as well as improved variability from HKMG technology. Overall, the industry was in a holding pattern, unable to scale the effective oxide thickness after the 90nm node until high-k came along, resulting in a reliance on strain techniques to boost speeds.

GlobalFoundries has a technology road map that relies on a gate-first HKMG for all of its low-power and high-performance offerings. Qualcomm’s decision to stick with a largely poly/SiON process raises the question of whether GlobalFoundries will switch gears and offer a poly SiON process at the 28nm node, in part to support Qualcomm, the largest foundry customer worldwide.

Clifford was asked if Qualcomm would use GlobalFoundries as a foundry supplier at the 28nm node, but he said he preferred not comment, saying that choice of foundries is still under discussion.

A GlobalFoundries spokesman said all of the foundry’s 28nm offerings on its “public roadmap” are based on gate-first HKMG. “However, we are offering a 28nm Poly/SiON technology at the specific request of certain customers with product applications that do not require the performance and leakage benefits provided by our HKMG technology,” he said. GlobalFoundries will not be enabling a design ecosystem around 28nm Poly/SiON.

He added that the HKMG ramp is “still on track and we are seeing significant customer traction. We expect HKMG to be the volume leader at 28nm in both low power mobile applications as well as high performance wired applications.” Multiple customer designs have been silicon-validated, and test chips are in prototyping at the company Fab 1, in Dresden, Germany, and are on the way to early risk production.

GlobalFoundries and TSMC have engaged in a marketing battle over the benefits of their competing approaches to HKMG: gate first for GlobalFoundries and gate last for TSMC. GlobalFoundries claims that its gate first HKMG process has a 10% to 15% cost advantage over the TSMC gate last process.

Yeap said that at the 22/20nm generation, Qualcomm plans to use HKMG for nearly all of its products.

Leave a Reply

(Note: This name will be displayed publicly)