A new way of arranging advanced computer components called memristors on a chip could enable them to be used for general computing, which could cut energy consumption by a factor of 100.

This would improve performance in low power environments such as smartphones or make for more efficient supercomputers, says a University of Michigan researcher.

This is the memristor array situated on a circuit board. Credit: Mohammed Zidan, Nanoelectronics group, University of Michigan.

“Historically, the semiconductor industry has improved performance by making devices faster. But although the processors and memories are very fast, they can’t be efficient because they have to wait for data to come in and out,” said Wei Lu, U-M professor of electrical and computer engineering and co-founder of memristor startup Crossbar Inc.

Memristors might be the answer. Named as a portmanteau of memory and resistor, they can be programmed to have different resistance states–meaning they store information as resistance levels. These circuit elements enable memory and processing in the same device, cutting out the data transfer bottleneck experienced by conventional computers in which the memory is separate from the processor.

However, unlike ordinary bits, which are 1 or 0, memristors can have resistances that are on a continuum. Some applications, such as computing that mimics the brain (neuromorphic), take advantage of the analog nature of memristors. But for ordinary computing, trying to differentiate among small variations in the current passing through a memristor device is not precise enough for numerical calculations.

Lu and his colleagues got around this problem by digitizing the current outputs–defining current ranges as specific bit values (i.e., 0 or 1). The team was also able to map large mathematical problems into smaller blocks within the array, improving the efficiency and flexibility of the system.

Computers with these new blocks, which the researchers call “memory-processing units,” could be particularly useful for implementing machine learning and artificial intelligence algorithms. They are also well suited to tasks that are based on matrix operations, such as simulations used for weather prediction. The simplest mathematical matrices, akin to tables with rows and columns of numbers, can map directly onto the grid of memristors.

Once the memristors are set to represent the numbers, operations that multiply and sum the rows and columns can be taken care of simultaneously, with a set of voltage pulses along the rows. The current measured at the end of each column contains the answers. A typical processor, in contrast, would have to read the value from each cell of the matrix, perform multiplication, and then sum up each column in series.

“We get the multiplication and addition in one step. It’s taken care of through physical laws. We don’t need to manually multiply and sum in a processor,” Lu said.

His team chose to solve partial differential equations as a test for a 32×32 memristor array–which Lu imagines as just one block of a future system. These equations, including those behind weather forecasting, underpin many problems science and engineering but are very challenging to solve. The difficulty comes from the complicated forms and multiple variables needed to model physical phenomena.

When solving partial differential equations exactly is impossible, solving them approximately can require supercomputers. These problems often involve very large matrices of data, so the memory-processor communication bottleneck is neatly solved with a memristor array. The equations Lu’s team used in their demonstration simulated a plasma reactor, such as those used for integrated circuit fabrication.

Researchers at the National Institute of Standards and Technology (NIST) have made a silicon chip that distributes optical signals precisely across a miniature brain-like grid, showcasing a potential new design for neural networks.

NIST’s grid-on-a-chip distributes light signals precisely, showcasing a potential new design for neural networks. The three-dimensional structure enables complex routing scheme, which are necessary to mimic the brain. Light could travel farther and faster than electrical signals. Credit: Chiles/NIST

The human brain has billions of neurons (nerve cells), each with thousands of connections to other neurons. Many computing research projects aim to emulate the brain by creating circuits of artificial neural networks. But conventional electronics, including the electrical wiring of semiconductor circuits, often impedes the extremely complex routing required for useful neural networks.

The NIST team proposes to use light instead of electricity as a signaling medium. Neural networks already have demonstrated remarkable power in solving complex problems, including rapid pattern recognition and data analysis. The use of light would eliminate interference due to electrical charge and the signals would travel faster and farther.

“Light’s advantages could improve the performance of neural nets for scientific data analysis such as searches for Earth-like planets and quantum information science, and accelerate the development of highly intuitive control systems for autonomous vehicles,” NIST physicist Jeff Chiles said.

A conventional computer processes information through algorithms, or human-coded rules. By contrast, a neural network relies on a network of connections among processing elements, or neurons, which can be trained to recognize certain patterns of stimuli. A neural or neuromorphic computer would consist of a large, complex system of neural networks.

Described in a new paper, the NIST chip overcomes a major challenge to the use of light signals by vertically stacking two layers of photonic waveguides–structures that confine light into narrow lines for routing optical signals, much as wires route electrical signals. This three-dimensional (3D) design enables complex routing schemes, which are necessary to mimic neural systems. Furthermore, this design can easily be extended to incorporate additional waveguiding layers when needed for more complex networks.

The stacked waveguides form a three-dimensional grid with 10 inputs or “upstream” neurons each connecting to 10 outputs or “downstream” neurons, for a total of 100 receivers. Fabricated on a silicon wafer, the waveguides are made of silicon nitride and are each 800 nanometers (nm) wide and 400 nm thick. Researchers created software to automatically generate signal routing, with adjustable levels of connectivity between the neurons.

Laser light was directed into the chip through an optical fiber. The goal was to route each input to every output group, following a selected distribution pattern for light intensity or power. Power levels represent the pattern and degree of connectivity in the circuit. The authors demonstrated two schemes for controlling output intensity: uniform (each output receives the same power) and a “bell curve” distribution (in which middle neurons receive the most power, while peripheral neurons receive less).

To evaluate the results, researchers made images of the output signals. All signals were focused through a microscope lens onto a semiconductor sensor and processed into image frames. This method allows many devices to be analyzed at the same time with high precision. The output was highly uniform, with low error rates, confirming precise power distribution.

“We’ve really done two things here,” Chiles said. “We’ve begun to use the third dimension to enable more optical connectivity, and we’ve developed a new measurement technique to rapidly characterize many devices in a photonic system. Both advances are crucial as we begin to scale up to massive optoelectronic neural systems.”