Behind The Intel-Altera Deal

The potential, and the pitfalls, to making this acquisition work.


Intel completed its $16.7 billion acquisition of Altera this week, wrapping up what is arguably the semiconductor industry’s most important M&A transaction of 2015. Time and numbers will tell exactly how important.

There are two big challenges to making this deal work. One involves a big shift in direction away from simply shrinking features to include new architectures and packaging approaches. If the idea works, it could have a significant impact on performance, power and customization options, not to mention the tools and equipment required to build new chips.

Intel has been a staunch proponent of Moore’s Law from the get-go, and for good reason. Three years after Gordon Moore penned his famous observation in a 1965 paper, Moore co-founded Intel. As you might expect, Intel has been one of the biggest, if not the biggest, beneficiary of Moore’s Law over the past five decades. It has perfected, with laser-like focus, the ability to get the next generation of processors out every couple years with a significant reduction in the cost per transistor.

Intel has kept its edge by investing at the forefront of process technology, and those continuous investments make it a very credible bellwether for just how hard it has become to remain on a two-year feature-shrinking cycle. In late 2014 the chipmaker began 14nm finFET production some six months after the expected target date due to yield issues. And it has pushed out its 10nm introduction until 2017, according to numerous reports. Even by revised Moore’s Law standards, shifting from an 18-month to a 24-month cadence, this is a noticeable delay.

Intel isn’t alone. The migration to production chips using 14nm and 10nm BEOL processes requires a series of technical feats that have never been attempted before. Pitches are severe, EUV lithography is still not ready, and the tools used for 3D metrology and inspection have been running out of steam for at least one or two process nodes. All of this adds up to higher costs due to longer times for development, manufacturing, and process optimization. Intel had hoped to drive down those costs with bigger wafer sizes—450mm versus the current 300mm—but that effort has stalled, maybe forever. Big equipment makers aren’t sure there are enough customers to warrant the investment, and foundries have balked at the challenges of physically handling and processing larger wafers.

This is no shock to companies working on the leading edge of technology, and all of the major server chip companies—Intel, IBM and AMD—have been working on alternative approaches for years. Intel began pushing systems on chip rather than just microprocessors while Paul Otellini was CEO. And it has been talking up the benefits of 2.5D and full 3D architectures based on through-silicon vias for at least the past several years, starting with the development of a low-cost interconnect between die called the Embedded Multi-die Interconnect Bridge. (AMD and IBM already offer 2.5D packages)

Those changes are very relevant to the Altera acquisition because FPGAs bring programmability into the processor world. If that programmability can be connected using high-speed interposers or TSVs, it could be dynamic rather than fixed (or nearly fixed), almost instantaneous — and a potential game-changer. While it’s possible to accomplish similar tasks in software, it’s much faster in hardware and it can be done much more efficiently. The ability to turn on and off logic to conserve power (and therefore reduce utility bills), as well as prioritize certain operations as needed, is long overdue in data centers. Facebook and Google have gravitated to their own architectures because of that, and losing other opportunities to more flexible architectures offered by ARM-based systems is a growing risk for Intel.

Intel has been lining up many pieces that can be layered into these systems for some time, including security software, real-time operating systems, and analytics. The side benefit is that it all could feed into Intel’s foundry business, where it controls the process, the packaging and the architecture.

But Altera is a sizeable company by itself, and digesting big companies with different cultures, processes and goals isn’t easy. That leads to the second big challenge: execution. Intel always has possessed some of the leading engineering talent in process, design, memory and even tooling. But its internal culture is so focused on improving the speed of the next round of processors and shrinking the size of everything inside of them, that it has come at the cost of ignoring other opportunities.

In the 1990s, the company made several acquisitions and forays into communications servers before abandoning the market. It completely missed the cellular phone and smart phone revolution. It even missed the FPGA market, selling off its programmable technology to Altera in the 1990s for a fraction of what it paid, basically buying it back with the current Altera purchase.

As noted in earlier stories, FPGAs are standard products, and Intel has been most successful with standard products. But it remains to be seen if Intel, or anyone else, can use standard products in non-standard ways to reach into new markets such as the Internet of Everything, where cost, power and time to market are critical, and where customization will be essential.

This requires, to paraphrase a well-worn marketing term, thinking well outside of the box. Intel’s history and culture, however, have focused on thinking deep inside the box, and then shrinking the box. Buying companies is a legal transaction. Leveraging those acquisitions in new ways is much harder. And it’s particularly difficult when it requires a substantial change in direction for the company making the acquisition, not to mention a potential change in the culture that still pays bills.


witeken says:

I’m not sure why the author insists Moore’s Law to have an 18 month cadence. It never has. It’s always been ~2 years since 2000. Which is just what Moore predicted in this follow-up in 1975.

Ed Sperling says:

Technically you’re correct, but former chairman and CEO David House predicted it would double every 18 months.

Ed Sperling says:

Technically you’re correct. But David House, former Intel chairman and CEO, put the number at 18 months during his tenure at Intel.

Leave a Reply

(Note: This name will be displayed publicly)