Chip-Level Advances That May Change Computing

Imagine a world with electronic devices that can power themselves, music players that hold a lifetime of songs, self-healing batteries, and chips that can change abilities on the fly. Based on what's going on in America's research laboratories, these things are not only possible, but likely.

Imagine a world with electronic devices that can power themselves, music players that hold a lifetime of songs, self-healing batteries, and chips that can change abilities on the fly. Based on what's going on in America's research laboratories, these things are not only possible, but likely.

"The next five years will be a very exciting time for electronics," says David Seiler, chief of the semiconductor electronics division at the Department of Commerce's National Institute of Standards and Technology (NIST) in Gaithersburg, Md. "There will be lots of things that today seem like far-out fantasy but will start to be commonplace."

In this two-part series, we'll take you on a tour of what could be the future of electronics. Some of these ideas may sound fantastic, others simply a long-overdue dose of common sense, but the common thread is that they have all been demonstrated in the lab and have the potential to become commercially available products in the next five years or so.

Today's story focuses on chip-level advances, from processors that transmit data via lasers instead of wires to circuits made with new materials that leave conventional silicon ones in the dust. These technologies could be the building blocks for a plethora of new and innovative products, some of which we can't even conceive of today.

Chips without wires: The laser connection

An up-close look at any microprocessor reveals millions of tiny wires going every which way to connect its active elements. Go below the surface, and there can be more than five times as many wires. Jurgen Michel, a researcher at MIT's Microphotonics Center in Cambridge, Mass., wants to replace all those wires with flashing germanium (Ge) lasers that transmit data via infrared light.

"As processors get more cores and components," explains Michel, "the interconnecting wires become clogged with data and are the weak link. We're using photons, rather than electrons, to do it better."

[ Related reading: Shining a light on the chip interconnect bottleneck ]

Capable of moving data at, well, the speed of light, a Ge laser can transmit bits and bytes 100 times faster than electricity moving through wires can, which means the critical connections between the chip's processing cores and its memory, for example, won't hold the rest of the device back. Just as fiber-optic communications made phone calls more efficient a generation ago, using lasers inside chips could put computing into overdrive.

The best part is that MIT's system doesn't require tiny fiber cables buried inside each processor. Instead, the chip is crisscrossed with a series of subterranean tunnels and caverns to transmit the pulses of light; tiny mirrors and sensors relay and interpret the data.

This mixing of traditional silicon electronics with optical components -- a practice known as silicon photonics -- can make computers greener as well. That's because lasers use less power than the wires they replace and give off less heat that needs to be cooled.

"Optoelectronics is a holy grail," says Seiler. "It can broaden electronics and is a great way to cut power use because you don't have all those wires acting like space heaters."

In February 2010, Michel and his collaborators, Lionel Kimerling and Jifeng Liu, successfully created and tested a functioning circuit that incorporates Ge laser data transfers. The chip hit speeds of over a terabit per second, or two orders of magnitude faster than today's best chips with wired connectors can do.

The chip is manufactured using current semiconductor processing techniques with some small additions, so Michel thinks that the transition to laser-based connections can happen over the next five years. If further tests are successful, MIT says it will license the process; this type of chip could become available around 2015.

The need has never been greater. By 2015, it's likely that there will be computer chips with up to 64 independent processing cores, each working simultaneously. "Connecting them with wires is a dead end," Michel says. "Using a germanium laser to connect them has huge potential and a big payoff."

Novel circuits: Memristors

If your MP3 player is filling up with tunes but you feel like a cultural murderer every time you delete a song, memristor technology might be arriving just in time.

The first fundamentally new electronic component to appear since silicon transistors came on the scene in the 1950s, the memristor presents a faster, more durable and potentially much cheaper alternative to flash memory. And with about twice the capacity of flash chips, there's plenty of room for everything from Leonard Bernstein to Lady Gaga.

"If we were redesigning computer technology today, we'd use memristor memory," says R. Stanley Williams, senior researcher and head of the Quantum Science Research (QSR) group at HP Labs in Palo Alto, Calif. "It's the fundamental structure for the future of electronics."

The memristor -- basically a resistor with memory -- was first posited in 1971 by University of California, Berkeley, professor Leon Chua, but HP Labs' memristor prototypes weren't publicly demonstrated until 2008.

To build its memristors, HP uses alternating layers of titanium dioxide and platinum; under an electron microscope they look like a series of long parallel ridges. Below the surface is a similar setup at a right angle, producing a grid-like array.

"Think of it as a series of cubes that are 2 to 3 nanometers (nm) on a side," says Williams. (A nanometer is one-billionth of a meter, roughly one ten-thousandth the thickness of a human hair.)

The key is that any two adjacent wires can be connected with an electrical switch beneath the surface, creating a memory cell. By adjusting the voltage applied to the cubes, scientists can open and close tiny electronic switches, storing data like traditional flash memory chips. (See IEEE Spectrum's excellent 6-Minute Memristor Guide for more details from Williams about how memristors work.)

Called ReRAM, for resistive random access memory, these chips can store roughly twice the data as flash chips but are more than 1,000 times faster than flash memory and could last for millions of rewrite cycles, compared with the 100,000 that flash memory is certified for. The bonus is that ReRAM's read and write speeds are comparable, while flash takes a lot longer to write data than to read it.

HP and South Korea's Hynix have teamed up to mass-produce ReRAM chips that could be used in a variety of small devices, such as media players that can hold terabytes of songs, videos and e-books. They expect to see the first products on the market sometime in 2013.

ReRAM can also replace dynamic RAM in computers. Because it's nonvolatile, ReRAM won't lose its contents when the system is turned off or loses power, as DRAM does. In fact, Williams thinks it could lead to an era of instant-on computing. Even when today's devices are merely put to sleep instead of being fully shut down, it takes anywhere from a few seconds to a minute for them to access stored data upon awakening. But with ReRAM devices, you'd be able to pick up where you left off instantaneously.

What's more, Williams says, it's possible to stack memristor arrays on top of each other within a single chip. This could create 3-D memory elements that better use the space within a chip. Rather than just using the surface of the chip, these memory elements can be built deep down into the chip, creating much more memory in the same physical volume.

"There's no fundamental limit to the number of layers we can produce," adds Williams. "We can get to petabit chips within about 10 years." That's a million gigabits of storage space, or enough to hold more than a year's worth of high-definition video on a chip the size of a fingernail.

"The first application for memristors will likely be memory," says NIST's Seiler, "but there's much more to it than that. There's a lot of potential beyond memory."

Further out on the digital horizon, perhaps 20 years or so, the technology could rewrite basic computer design. In 2010, the HP researchers discovered that memristors can be used for logic computations in addition to storage. That means that both functions could theoretically occur in the same chip.

Says Williams, "A single memristor can replace a handful of other circuits, simplifying how computers are designed, made and operated." For instance, he says, one memristor could do the job of the six transistors that are currently used to create a single static RAM memory cell in a processor's cache.

Williams thinks it might even be possible to create an artificial neural synapse with memristor technology that could mimic the way the brain works. That's decades off in the future, however, if possible at all.

The memristor certainly has the power to rewrite the rules of electronics, says Supratik Guha, director of the physical sciences department at IBM. But, Guha notes, the technology still needs to prove itself. "There may be potential as a memory element," he says. "But like any other technology, you need to crawl before you walk and walk before you run."

In other words, memristor technology won't happen overnight. It will take a lot of evolution and time before memristors are as prevalent as DRAM or flash memory.

Changeable chips: Programmable layers

From the fastest processor to the smallest memory module, just about every chip used in electronics today has one thing in common: Its active elements reside in the top 1% to 2% of the silicon material it's made of.

That will change over the next few years as chip makers go vertical to squeeze in more and more components. Some vendors, such as Intel, have resorted to gluing completed chips together, while University of Rochester researchers have designed and built 3-D circuits layer by layer internally. Both approaches are enormously complicated and expensive, observes IBM's Guha.

But what if you could trick the circuit into rearranging itself on demand so that it only appeared to other components to have several layers of active elements? That's the idea behind Tabula's Spacetime technology and its ABAX chip design.

Rather than having several layers of hardwired components that are permanently etched into silicon and never change, ABAX uses reprogrammable circuits that can change their abilities on demand. Its current products deliver the equivalent of up to eight different chip layers that can be changed faster than the blink of an eye.

"Think of it as like a department store with eight floors," says Steve Tieg, Tabula's president and chief technology officer. "You'd take an elevator to go between floors to shop for different items." But rather than having eight different physical floors, each with its own internal arrangement and assortment of goods, Tabula has figured out a way to have a single layer (or floor) that reconfigures itself as needed.

"It's as if while you're on the elevator, they're inside rearranging the floor to create a different layout with different products," adds Tieg. "It looks to the outside world as if there are eight floors, but there's only one."

Glossary

Epitaxy: The process of growing a thin crystalline layer on the surface of another crystal so that the layer mimics the structure of the substrate.

Germanium: In between gallium and arsenic and one row below silicon on the periodic table of the elements, germanium (Ge) is used in fiber-optic cables.

Graphene: A single layer of carbon arranged in a honeycomb lattice, graphene has a variety of novel electronic properties, such as the electrons' ability to move much faster than in silicon.

Memristor: A novel electronic structure that combines memory with a resistor, this device can simplify and speed up how electronics perform.

Moore's Law: Postulated by Intel co-founder Gordon Moore in the late 1960s, it states that every two years the density of transistors on a chip will roughly double, making for ever more powerful chips.

ReRAM: Resistive random access memory is built from memristor technology and can replace flash memory.

To work, the chip's reprogrammable circuits are fed with its next series of assignments and duties in just 80 picoseconds -- 1,000 times faster than the chip's computational cycle. That way, the layers can be changed on the fly while the chip is waiting for its next commands.

In a real sense, ABAX does more with less. Made with conventional semiconductor processing technology, Tabula's ABAX chips cost roughly the same to manufacture as traditional ones. The design still uses only the top surface of the chip, but that single layer does the work of eight different chips. According to Tieg, the technology can increase the density of circuits twofold, and memory and video throughput can be boosted by as much as 3.5 times.

This idea could potentially usher in a new era in semiconductors where a single chip replaces several or adds abilities without the cost and power use of extra parts. "Virtualizing the operations of a chip can have a big payoff in terms of efficiency and flexibility," says NIST's Seiler. "The key is how it gets programmed."

Related:
1 2 Page 1
Page 1 of 2
Survey says! Share your insights in our 19th annual State of the CIO study