COSTA MESA, Calif. — Researchers at Irvine Sensors Corp. are ramping up for a new level of wiring density that could approach the interconnectivity of neurons in the brain.

With new funding in hand from the government's Small Business Innovative Research grant program, the company plans to leverage its chip-stacking technology by fully interconnecting the technology's various silicon layers, which would represent a major step toward its goal.

"I think we are between five and 10 years from having something with human-level computing power: 10 petaflops in less than 1/3 cubic foot and consuming less than 10 watts," said John Carson, chief technical officer of Irvine Sensors. Carson has been working for years toward building a silicon brain. While his engineers concentrate on the task at hand, he is looking past the current project toward the next development steps.

It's not hard to see why. If Irvine Sensors can squeeze a million analog neurons onto each chip, and achieve a thousand chips in a one-inch cube, then the arithmetic starts to get interesting. A thousand of these cubes, which could fit in a shoebox, would produce a system with a trillion neurons and two trillion interconnections. Although not quite on the same scale as the interconnectivity of the brain, that level of interconnect density would be a large step in the right direction. Carson's already looking to optoelectronics to connect the subsystems and to robotics to give his brain a body.

Despite the technical challenges in producing such a system, the time is right for this kind of push, said Dwight Duston, assistant deputy for technology at the U.S. Army Ballistic Missile Defense Organization (BMDO). "BMDO has been sponsoring research in all these [analog neural networks, chip stacking, photonics] technologies for many years now. We assessed the maturity of each individual technology and determined that the time was right to integrate them into an advanced module for testing and evaluation."

Despite some common operations, computers and brains are very different systems. Brains consist of a trillion (10¹²) or so neurons that act as both processor and memory. On top of that, there are a thousand trillion (10¹⁵) synapses that connect the network of neurons, allowing the brain to act as a single system. Today's microprocessors, on the other hand, have just a few million logic gates to process information. The data to be processed travels via a few hundred or a thousand in/out ports, which also act as bottlenecks. Though hundreds or thousands of individual chips can be connected to form a supercomputer, communication between chips is relatively slow and difficult. The effect is still powerful in a computational sense, but hardly brainlike.

So while we humans are good at making sense of lots of disparate information, the supercomputer must be reserved for large problems that can easily be split into independent, processor-intensive chunks.

Despite these problems, engineers have been using computers in brainlike ways for years, configuring them as artificial neural networks. These networks have successfully been used to learn to play board games, distinguish cancerous cells from healthy ones, determine who's creditworthy, and even predict the behavior of the stock market. However, their functionality has always been seriously limited by their size.

Irvine Sensors' core technology is a method for layering chips that crams 50 chips into the volume of a single large chip. The approach grinds away the silicon wafer, producing a structure as thin as paper. These chips are then stacked and bonded to form a single unit.

The stacking technique, apart from its ability to get more devices into the same volume, also attacks the density-of-interconnect aspect so crucial to neural networks. In a conventional system, network-intensive architectures cannot expand to another processor because of the bottleneck caused by the I/O pins. In an ordinary stacked system, metal buses that run along the stack edge fulfill the same function as pins, and offer no more connectivity.

To come up with a solution to this interconnect problem, Irvine Sensors turned to the Army's missile-defense organization. Funding from the BMDO, in the form of a Phase I SBIR award, helped company engineers to invent the three-dimensional field-effect transistor. This device is constructed as two parts on two separate chips.

Once these have been bonded together in a chip stack, they operate as a single transistor. The invention allows a signal to be passed easily from one neuron to its neighbor, even if the two were on different chips.

Ultrahigh density
Though the device is now being patented, and has been demonstrated, it has not yet been proven in real systems. BMDO recently announced a Phase II SBIR award worth $738,000 to do just that. In the next 18 months, the goal is to develop the ultrahigh-density interconnect to the stage where it can unite a whole chip stack. Carson said the long-term goal is to integrate 1,000 neural chips in a single cube. This would involve thinning each chip down to 10 micrometers.

To make something useful out of all this hardware, Irvine Sensors has picked up two major partners: the Jet Propulsion Laboratory and the California Institute of Technology, both in Pasadena. In particular, JPL has done much of the neural design for the chip-stack analog neural networks that the company has already demonstrated. These systems, which consist of a camera connected directly to a neural network, act as very fast parallel image-processing and recognition devices.

At Cal Tech, researchers have been looking at the problem from a biological perspective. Professor Demetri Psaltis, head of the Neuromorphic Systems Engineering Center, has been working with his students to develop a model of the "V1" visual cortex.

Besides trying to build an optical implementation of the model in their own labs, the researchers are now working with Irvine Sensors to build it in silicon. This implementation, to be built under the SBIR, will connect a 3-D stack of field-programmable gate arrays.

Duston said the BMDO needs "a processor which is small, lightweight, generates little heat when operating and can process imaging data from many sensors with differing formats rapidly."

He said he sees wider implications as well. "For the past few decades," he said, "the biological and semiconductor communities have been working diligently to understand the processing mechanism of the brain, each approaching it from their own discipline and perspective. This is a major step toward that meeting point in the middle where we will finally understand the magical architecture of the brain."

In March of this year, the Boeing Co. gave Irvine Sensors a $1.3 million military subcontract to build a wearable voice-activated computer for soldiers, each system the size of a deck of cards. Irvine Sensors is also using its stacking technique in commercial products, including a four-layer flash-memory chip that can store 128 Mbits of data in a thin 0.5-inch by 0.5-inch square.