As the cost of making computer chips increases by leaps and bounds — a semiconductor factory or “fab” cost $2-3 billion to outfit in 2005; today, a new plant that can make 14nm chips will run you $10-12 billion — industry participants are twisting and turning every which way as they try to extend Moore’s Law yet further.
Not a law in the immutable physical sense, Moore’s Law was an observation by Intel’s Gordon Moore in 1965 that semiconductor technology seemed to be on a many-year path of doubling the number of transistors for a given area every 24 months. Moore himself re-upped on his law every few years as the silicon train seemed to remain well on track.
There was a false step in 2004, when players saw that just increasing clock frequency would no longer do the job. The free part of the developmental trajectory was over. After that, chip makers turned to the complex charms of parallelism — which in theory could deliver more work per time by breaking tasks into pieces, having separate engines work on them all at once, and reassembling the result at the end — but the programming model for parallel computing became a kind of black art in itself.
The way forward was no longer clear. And the costs of overcoming each new hurdle were rising. Since light is used to etch features on silicon slabs in chemical baths, companies began to use shorter wavelengths to carve smaller features. Immersing the process in purified water also helped refine control. But the machines to perform this increasingly exotic work continued to rise in complexity and price. ASML, a Dutch supplier of such systems, will sell you a Twinscan NXE:3300B production extreme ultraviolet (EUV) lithography machine for about $100 million.
More recently, first
Which brings us to 14nm, the next commercial process. Not counting the fabless-semi companies — those, like
In 2012, at a conference in Santa Clara, CA, IBM scientists first began sowing the seeds of doubt about future-generation technology. Gary Patton, then a vice president at IBM’s semiconductor unit, told the crowd that he wasn’t sure 10nm was worthwhile. That is, he seemed to be saying while getting to 10nm might be possible, it would be so expensive that the benefits of smaller features might not justify the investment. He said beyond that, at 7nm, all bets were off. Any features smaller than 12 atoms in thickness would run into Heisenbergian instability and could not stably hold data. A lot of eyebrows went up.
Intel, on the other hand, doubled down, saying it expected to continue its march toward Moore’s ever-receding horizon. Both companies (and others) began to talk about alternative materials and technologies, such as carbon nanotubes, graphene structures, and other exotic projects. Development continues in these areas.
Meanwhile, however, there’s the matter of the next few years, when all firms are expected to take silicon technologies down through 14nm, 10nm, and maybe 7nm using existing tricks and a whole bag of new ones. The cost of this arms race, however, will continue to rise.
Like a great poker player, Intel just sits at the table, raising the pot each hand. Intel executives continue to exude confidence in the company’s superiority in process node technology, and it’s hard to tell whether they’re bluffing about the out years. IBM is hedging its bets.
One of the many ways IBM is seeking to defray future development costs is through a potential spinout of silicon manufacturing assets. Heavy sideline betting says the result of this rebirth will be a new entity in which IBM — which needs to keep the confidence of its silicon customers in its commitment to long-term development — retains an minority ownership interest. The company recently reinforced this message with the announcement of a $3 billion program to both refine existing processes and develop new ones.
Another means by which IBM is spreading some of the risk associated with semiconductor development is its donation of key technology in its Power architecture to a consortium called OpenPOWER. For the past couple of years, Power architecture has been the basis for several IBM server lines and not much else. Before that,
Interestingly, 11 of these 56 companies are Chinese. You may be aware that the Chinese government and industry are objecting increasingly vociferously to the high rents they have to pay in the form of license fees and elevated product prices to foreign companies like Qualcomm and
At a recent analyst conference in Greenwich, CT, one of the Chinese OpenPOWER members, Suzhou Powercore Technology’s Lixin Zhang, said during a panel, “There is a saying in China: Intel takes the meat and leaves you just the bones." Suzhou, based in Beijing, is developing servers with Power technology.
It’s clear that the rebellious attitude of large swaths of the Chinese political economy toward foreign holders of key intellectual property is playing out in IBM’s favor, a factor that may help the company keep the glow of value alive in its Power architecture as the cost of staying in the semiconductor game increases for everyone.
Disclosure: Endpoint has a consulting relationship with several of the companies mentioned in this article.
