Intel has long led the way, but as it gets increasingly harder for the industry to follow Moore's Law, there has been a lot of debate about whether its lead is growing too large. With wireless customers clamoring for the latest technology for mobile processors, semiconductor foundries are suddenly shaking things up.
With each new generation of processors, as feature sizes approach atomic scale, it is getting harder--and more expensive—for chipmakers to follow Moore’s Law. Intel has long led the way, but lately there has been a lot of debate about whether its lead is growing a bit too large. In a bid to catch up, the rest of the industry is suddenly changing its roadmaps and forging new alliances.
Intel has deep pockets, and because it makes its own chips, it can tailor the manufacturing process and processor design to get the best results. That means it is first to release chips with the smallest dimensions (not counting memory, which is a different animal). These “shrinks” worked for decades. But starting around 2002, when Intel released its first 130-nanometer Pentium 4 processors, simple scaling began to run out of steam. Since then Intel has introduced a series of innovations--strained silicon, new gate materials, and most recently, 3D transistors--to continue to make chips that are denser, faster and use less power. Intel is now counting on this technology lead, not only to defend its monopoly in PCs and servers, but also to expand into smartphones and tablets.
At one time there were lots of other IDMs (Integrated Device Manufacturers), but an increasing number of companies design chips and hire foundries to manufacture them. AMD is the most prominent example, having spun-off its manufacturing arm to establish GlobalFoundries, but many others have turned to fab-lite or fabless strategies. Major semiconductor companies such as Freescale, NXP Semiconductors, STMicroelectronics, Texas Instruments and Toshiba have all shifted more work to foundries in recent years. That cuts costs, but it also leaves them at the mercy of foundries. This works fine as long as the foundries stay close behind Intel. But as scaling becomes more difficult, the foundries have struggled to duplicate innovations like high-k materials and metal gates (HKMG) or 3D transistors, better-known as FinFETs. The foundries have the added challenge that they must serve lots of different customers and build lots of different types of products.
Intel began volume production of 22nm processors using FinFETs late last year. On the company’s recently quarterly call, CEO Paul Otellini said the 22nm “ramp” was ahead of schedule and the latest Ivy Bridge chips are now a quarter of its total PC production. Intel doesn’t provide exact numbers, of course, but that means it is churning out perhaps 225,000 Ivy Bridge chips per day on its 22nm process with FinFETs. This will be followed by a new microarchitecture, code-named Haswell, early next year, and by early 2014 Intel is scheduled to ship its first 14nm processors.
Meanwhile the foundries are ramping the 32nm/28nm node and perfecting HKMG technology. TSMC, the world’s largest foundry, shipped its first mainstream 28nm product, AMD’s Radeon HD 7970 graphics processor, in December 2011, but it won’t be able to meet demand for 28nm capacity until the end of this year. After struggling with manufacturing yields last year, GlobalFoundries now seems to be producing enough AMD processors on its 32nm process with HKMG, but it won’t get to 28nm until next year. Similarly UMC won’t have 28nm with HKMG in volume production until next year. Samsung’s latest Exynos 4 Quad processor is manufactured on a 32nm process, with HKMG, but it still manufactures Apple’s A5X, the chip in Apple’s latest iPad, on an older 45nm process. Until recently, no foundry had planned to shift to FinFETs until the 14nm node, which is still years away.
This gap was no big deal as long as the demand for the most leading-edge technology was relatively limited. TSMC made a small batch of chips—typically programmable logic for Altera or Xilinx or high-end GPUs for AMD or Nvidia—on the latest node, and then took time to fine-tune the process before opening it up to other customers. Other foundries came along a year or so later and competed largely on price. But the mobile revolution has changed all of that. Now customers such as Apple, Qualcomm, Texas Instruments and Broadcom are leading the charge. They want application processors and basebands on the latest process technology in high volumes to meet demand for smartphones and tablets. To keep pace the foundries are suddenly shaking things up.
Last week ARM and TSMC announced that they will work together to develop chips that will use a 16nm process with FinFETs. TSMC said it hoped to have it in volume production sometime in 2015. These chips will also have cores based on the upcoming ARMv8, ARM’s first 64-bit architecture, which means devices will be able to a make use of more than 4GB of memory. That will be useful for high-end smartphones and tablets a few years down the road, but it’s critical for ARM-based PCs and servers. TSMC is also adding lots of new capacity and has talked publicly about building fabs devoted to large customers, much in the same way that Samsung’s fab in Austin, Texas is devoted to Apple.
Rival UMC announced last month that it had licensed IBM’s 20nm technology with FinFETs and plans to offer it to customers as early as 2015. UMC is also shopping a 10 percent stake in the foundry hoping to lure a strategic partner or large customer such as Qualcomm that can help it keep up with the competition.
GlobalFoundries and Samsung are part of IBM’s joint development alliance, so they have access to the same technology. So far GlobalFoundries still plans to use conventional transistors at 20nm and introduce FinFETs at 14nm in early 2015. Samsung seems to be on a similar schedule—it has previewed early 20nm wafers with planar transistors and 14nm wafers with FinFETs at industry conferences. Both companies are adding lots of new capacity too.
In some ways Intel’s lead isn’t quite as big as it seems. ARM CEO Warren East told EE Times last week that the real competition isn’t Intel’s PC processor, but rather its Atom SoC (System on Chip), which is still based on an older 32nm process with standard planar transistors. That’s true, but Intel has announced plans to speed up SoC development. It plans to release a 22nm version next year followed by 14nm in 2014, perhaps with its Infineon baseband technology integrated on the same chip. Intel is making a little progress in smartphones, with models from Lenovo, Orange and Lava available in some markets, but if it can execute on this plan by 2014 it should be a much stronger competitor in smartphones and other mobile devices.
Conversely the ARM camp has designs on PCs and servers. The first Windows RT devices will go on sale when Windows 8 launches on October 26. This will likely include Microsoft’s own Surface tablet, the Asus Tablet 600 and perhaps a Lenovo tablet all based on Nvidia’s Tegra 3; a Samsung device based on a Qualcomm chip; and a Toshiba tablet using TI’s OMAP4. These will compete directly with x86-based tablets and convertibles running Windows 8. Longer term, several chipmakers including AMD (Sea Micro), AppliedMicro, Calxeda and Marvell are developing ARM-based processors for servers where Intel’s Xeon dominates the market.
The gap between Intel’s process technology and the rest of the industry will be one of the key factors in this battle over the next few years.