Russia has announced a breakthrough in a key component for next-generation artificial intelligence hardware. A new type of memristor has been developed by researchers at the Moscow Institute of Physics and Technology (MIPT) to address an ongoing difficulty in devices designed to emulate the information-processing capabilities of the human brain. This memristor is capable of retaining information for almost 12 days.
The development has the potential to advance neuromorphic computing, an emerging approach that aims to integrate memory and computation in hardware, thereby potentially reducing the energy consumption associated with conventional AI processors. The Russian researchers have also reported a significant increase in the device’s rewriting endurance, which is important to the development of artificial synapses that are more enduring.
The Operation of the Russian Memristor
A memristor is an electronic component that undergoes an electrical resistance change in response to the voltage applied to it. The device can execute functions associated with both memory and computation using its resistance to represent a stored state.
This characteristic renders memristors an appealing option for neuromorphic processors. Neurons are connected by synapses in the human brain, which alter their strength as learning occurs. By adjusting its electrical resistance, a memristor can effectively represent a varying synaptic weight, thereby imitating aspects of this behavior.
Energy and time are typically consumed by conventional computing systems as they transfer data between a processor and separate memory. Memristor-based architectures are designed to minimize this movement by enabling computation to occur in close proximity to or directly within the elements that store information.
Nevertheless, the construction of a dependable artificial synapse necessitates a complex engineering trade-off. Devices that are able to learn quickly and respond readily to electrical signals may encounter difficulty in maintaining their configured states. In contrast, devices that are intended for storage stability may be more difficult to modify during the learning process.
The MIPT announcement indicated that earlier devices encountered challenges in retaining information for an extended period. This limitation was resolved by the Russian team, which investigated the fundamental physical processes that regulate resistance changes.
Semiconductor Electrodes and an Optimised Thin Film
The researchers modified the device’s structure by replacing metallic electrodes with semiconductor electrodes. Additionally, they determined the optimal thickness for the intermediate film, which was composed of zirconium and hafnium compounds.
The memristor was reportedly able to operate in two distinct modes as a result of these modifications: one that was designed to adapt its state during learning and another that was designed to preserve information for an extended period.
The resulting improvement is significant. The results of the research reported indicate that the number of rewriting cycles increased by approximately twentyfold, while the information-retention time increased from several seconds to almost 12 days.
These figures describe the performance of the reported device, not a finished commercial processor. Its practical importance will be contingent upon supplementary measurements, such as energy consumption, switching speed, device-to-device variation, operating temperature, manufacturing yield, and long-term reliability.
However, the improvement of both retention and rewriting endurance addresses two critical requirements for hardware that must continuously learn while simultaneously preserving valuable information.
Which nations are currently developing comparable technologies?
Universities, semiconductor manufacturers, and specialist companies are developing memristors, resistive memory, and computing architectures inspired by the brain in an international research field into which Russia is now entering. Nevertheless, not all neuromorphic hardware products use memristors, and not all resistive memory technologies replicate the learning behavior exhibited by the MIPT device.
There are many international examples that serve as illustrations of the competitive environment.
United States: Crossbar and Knowm
Knowm is a company headquartered in the United States that specializes in the development of memristors and crossbar arrays for the purpose of conducting research on artificial synapses and machine learning. Its technology offers a pertinent comparison, as researchers can acquire physical memristor components for experimentation, rather than relying solely on computer simulations.
Resistive random-access memory (ReRAM) technology is developed by Crossbar, an additional United States-based organization. ReRAM targets embedded computing applications and non-volatile memory by utilizing variations in electrical resistance to store information. Although it is associated with memristor research, its commercial objectives are distinct from MIPT’s emphasis on the preservation of adjustable synaptic states.
Additionally, the United States maintains a considerable presence in neuromorphic computing via Intel’s Hala Point research system and Loihi processors. These systems employ brain-inspired computational architectures; however, they are not direct equivalents of the Russian memristor due to their distinct hardware approach.
IBM Research has also explored analog in-memory computing, which involves the use of memory devices and circuits to expedite AI calculations. This research is indicative of the overarching goal of decreasing the energy and time necessary to transfer data between distinct storage and processing units.
Samsung Electronics, South Korea
Samsung Electronics is one of the foremost semiconductor companies that are involved in the research and development of resistive memory. The same overarching objective of integrating compact storage with resistance-based switching is pertinent to ReRAM.
Nevertheless, the retention periods, endurance characteristics, and applications of each implementation are distinct. The existence of commercial memory technologies does not necessarily imply that they can replicate the adaptive behavior necessary for neuromorphic systems.
Japan: Panasonic
Panasonic has conducted research on the integration of oxide-based ReRAM into semiconductor products. This is a distinct level of technological sophistication from a laboratory memristor that is intended to replicate a synapse. Embedded memory is designed for practical electronic systems, whereas neuromorphic computing necessitates additional capabilities to facilitate adaptive information processing.
Industrial Development and Research in China and Europe
China has established a significant research ecosystem that encompasses neuromorphic computing, resistive memory, and memristors. In order to facilitate energy-efficient artificial intelligence, universities and research institutes have examined a variety of materials, device structures, and architectures.
Other European research groups have also made contributions to the development of memristors, analog computing, and brain-inspired hardware. Alternative device materials and methods of integrating memory with computation are the subjects of research programs throughout the continent.
Australia and other nations are also involved in the broader resistive-memory research ecosystem. However, laboratory research, experimental production, and commercial manufacturing should not be considered equivalent accomplishments.
How Many Countries Make Memristors?
There is no universally accepted figure for the number of countries manufacturing memristors. The response is contingent upon whether manufacturing refers to the production of research components, laboratory fabrication, commercial embedded-memory devices, or the mass production of complete neuromorphic processors.
Memristive devices and related technologies are the subject of research or industrial development in at least six countries or major national research ecosystems: the United States, Russia, China, Japan, South Korea, and various European countries. This field is also supported by Australia and other nations.
This is an illustrative list, not a verified count, of the countries that manufacture commercially available memristors on an industrial scale. A nation that conducts university research on memristors does not necessarily have a domestic manufacturer that supplies them in substantial quantities.
The industry remains at different stages of development. Some resistive-memory technologies are currently being developed for commercial semiconductor applications, while more sophisticated artificial synapses and neuromorphic architectures continue to be the subject of research and engineering validation.
Can Russia Turn This into an AI Processor?
The MIPT result is promising because it addresses a fundamental deficiency of artificial synapses: the capacity to learn without rapidly forgetting what they have learned. These components could be more beneficial in systems that must update information repeatedly while preserving previously acquired states due to their increased rewriting endurance and longer retention.
Integration represents the next obstacle. It is critical that researchers demonstrate the consistent operation of large arrays of these devices, the precise programming of their states, and the overall system’s measurable improvements in performance and energy consumption.