IBM will invest $3 billion over five years to develop processors with much smaller, more tightly packed electronics than today's models, and to sustain computing progress even after today's manufacturing technology runs out of steam.
The first goal is to build chips whose electronic components, called transistors, have features measuring just 7 nanometers, the company announced Wednesday. For comparison, that distance is about a thousandth the width of a human hair, a tenth the width of a virus particle, or the width of 16 potassium atoms side by side.
The second goal is to choose among a range of more radical departures from today's silicon chip technology -- a monumental engineering challenge necessary to sustain progress in the computing industry. Among the options are carbon nanotubes and graphene; silicon photonics; quantum computing; brainlike architectures; and silicon substitutes that could run faster even if components aren't smaller.
"In the next 10 years, we believe there will be fundamentally new systems that are much more efficient at solving problems or solving problems that are unsolvable today," T.C. Chen, IBM Research's vice president of science and technology, told CNET.
Scientists and engineers have postponed the transition to this "postsilicon" future many times, but atomic-level constraints eventually will block today's basic manufacturing approach. Already quantum-physics problems like electrons "tunneling" from one place to another complicate chip design.
Moore's Law, an observation by Intel co-founder Gordon Moore that the number of transistors on a chip doubles every two years, seems almost a given in the computing industry. But it takes sustained work in research and development, and it's not easy. Even the industry leader, Intel, has troubles: last year, it delayed the debut of its 14nm "Broadwell" chip designs from 2013 until 2014 because of a defect issue.
That steady progress is tremendously important. Cramming more transistors into a given surface area means mobile phones can play video instead of just make phone calls, watches can notify you of appointments instead of just tell time, and authentication chips can be built into credit cards. If Moore's Law fizzled, the arrival of everything from supercheap smartphones to Google's world domination could be slowed.
The 7nm technology is three generations of manufacturing into the future from IBM's current flagship processor, the Power8, built with a 22nm process. IBM's research and development work will focus on finding the materials and processes necessary to make 7nm chips economically viable, the company said.
"The question is not if we will introduce 7-nanometer technology into manufacturing, but rather how, when, and at what cost," said John Kelly, senior vice president of IBM Research, in a statement.
Big Blue's money will fund teams at labs in New York, California, and Switzerland, and the company will hire new staff for the work.
On the road to 7nm chips will be stops at 14nm and 10nm. The map gets fuzzier after 7nm. Intel expects at least one more step, the 5nm manufacturing process, and chip manufacturing toolmaker Applied Materials has discussed 3nm after that (PDF). But IBM didn't specify what exactly it thinks will come after 7nm.
IBM is a leader in materials science, chemistry, physics, and nanotechnology, and it boasts of having twice the patents on postsilicon processor approaches as any competitor. However, it hasn't matched the manufacturing volume of today's powerhouses: Intel, Samsung, and TSMC.
IBM justifies its chip business with a higher-level mandate, though. It doesn't just make chips and sell them, it makes chips and uses them in computers that serve other priorities -- big-data analysis machines for corporate customers and supercomputers to simulate the human brain for government-funded research, for example. And major changes in manufacturing technology could reorder the current power structure.
IBM's new research will investigate several avenues beyond today's manufacturing technology:
- Neurosynaptic computing: IBM is working on technology that moves away from today's decades-old one-step-after-another approach to a more brainlike design that relies on a computing equivalent of brain cells, called neurons, and their electrical communication pathways, called synapses. IBM ultimately hopes to make a system with 10 billion neurons operating in parallel using 100 trillion synapses, consuming less than a kilowatt of power, and occupying less than two liters of volume.
- Quantum computing: This unusual technology relies on the shift from today's digital, using bits that represent either a zero or a one. Instead, quantum computing uses qubits that can hold both those values at once, which means they could be used in principle to perform many calculations at the same time. Some experimental quantum computers have arrived on the market -- Google and NASA have one $15 million model from D-Wave -- but they're still very much an unknown quantity.
- III-V materials: Today's chips usually use silicon, which is a member of group IV of the periodic table of the elements. Faster chips can be built using elements from groups III and V -- gallium arsenide, for example. They run faster because electron mobility is higher, which determines how fast the transistors can switch on and off. However, today they're not economical for mainstream electronics. IBM thinks they could be, and for lowering the power consumption of computing.
- Carbon nanotubes: Carbon atoms can bind together into a tubular structure called a carbon nanotube (CNT). For IBM's post-7nm chips, CNTs are a candidate for replacing silicon in the role of chip semiconductor -- a material that either conducts electricity or not, depending on external circumstances. Semiconductors are the core part of chip transistors, which act like tiny on-off switches. Nanotubes might be usable for much smaller transistors, and their fast switching speeds could mean a fivefold or tenfold performance boost over silicon.
- Graphene: A close relative of carbon nanotubes is graphene, a flat lattice of carbon atoms just one atom thick. Because electrons move fast in graphene, IBM hopes the material could be used for quick-response tasks like managing high-frequency radio signals.
- Tunnel field effect transistors: Tunneling is a problem today, but an approach called tunnel field effect transistors (TFETs) uses the phenomenon to drive a transistor's electrical current. A voltage difference across chip elements drives electrical current in today's chips, but these "steep slope" devices could require a lower voltage difference and therefore reduce the waste heat that holds chips back. IBM thinks TFETs can cut power consumption by a factor of 100.
- Silicon photonics: Photons -- light particles -- travel faster than electrons, and they don't produce the debilitating waste heat of electricity flowing through wires. That's the core advantage of silicon photonics, which uses chips to generate and receive signals. An added bonus: you can transmit multiple frequencies of light, packing more information into a signal. Fiber-optic networking is in widespread use today, but it's economical only over long distances; IBM is among those working to shorten links to connect computers within data centers and later connect components within a single computer.
IBM wouldn't comment on when or how it plans to commercialize the new technology, which is still in the hands of its labs, not its product design or manufacturing group.
But it's clearly feeling the time pressure.
"We anticipate that in order to scale to 7nm and perhaps below for the industry, we will need to have the semiconductor architectures and new manufacturing tools and techniques in place by the end of the decade," Chen said. "That's why it is critical for us to make the significant investment now into the research and early-stage development to demonstrate what 7nm innovations will be useful, before it can even be commercialized."