NEURRAM

Company creates neuromorphic processor for AI

The NeuRRAM chip is the first in-memory computing chip that demonstrates a wide range of AI applications using only a small percentage of the power consumed by other platforms while maintaining equivalent accuracy.

NeuRRAM, a new chip that performs calculations directly in memory and can run a wide range of AI applications, was designed and built by an international team of researchers. What sets it apart is that it does all this for a fraction of the power consumed by general-purpose AI computing platforms.

The neuromorphic NeuRRAM chip brings AI closer to running on a wide variety of peripherals disconnected from the cloud. This means they can perform complex cognitive tasks anywhere and anytime without relying on a network connection to a centralized server. Applications for this device are plentiful in every corner of the globe and in every aspect of our lives. They range from smartwatches to virtual reality headsets, smart headsets, smart sensors in factories and rovers for planetary exploration.

The NeuRRAM chip is not only twice as energy efficient as current in-memory computing chips, an innovative class of hybrid chips that perform in-memory computing, but also delivers results as accurate as digital chips. conventional. Conventional AI platforms are much larger and often limited to using large data servers running in the cloud.

Furthermore, the NeuRRAM chip is very versatile and supports many different neural network models and architectures. As a result, the chip can be used for many different applications, including image recognition and reconstruction and speech recognition.

It's often thought that greater memory computing efficiency comes at the cost of versatility, but our NeuRRAM chip offers efficiency without sacrificing versatility,” said Weyer Wang, first author of the paper.

Currently, AI computing is energy intensive and computationally expensive. Most AI applications on edge devices involve moving data from devices to the cloud, where AI processes and analyzes it. Results are then transferred back to the device. This is necessary because most peripherals are battery powered and, as a result, only have a limited amount of power available for computing.

By reducing the power consumption required for AI inference at the edge, the NeuRRAM chip can lead to more reliable, smart and affordable edge devices and smarter manufacturing.

It can also lead to greater data privacy as transferring data from devices to the cloud carries greater security risks.

In AI chips, moving data from memory to computing devices is one of the main bottlenecks. That equates to an eight-hour commute for a two-hour workday,” said Weier Wang.

To solve this problem with data transfer, the researchers used so-called resistive random access memory. This type of non-volatile memory allows calculations to be performed directly in memory, rather than in separate computing units.

Chip performance:

-The researchers measured the chip's energy efficiency on a metric known as the energy-delay product (EDP).

-EDP ​​combines both the amount of energy consumed for each operation and the time required to complete the operation. According to this indicator, the NeuRRAM chip provides a 1.6 to 2.3 times reduction in EDP (the less the better) and a computational density 7 to 13 times greater than that of modern chips.

-Engineers performed various AI tasks on one chip. It achieved 99% accuracy in the handwritten digit recognition task; 85.7% in the image classification task; and 84.7% in the Google voice command recognition task.

In addition, the chip also achieved a 70% reduction in image reconstruction errors in the image restoration task. These results are comparable to existing digital chips, which perform calculations with the same precision but with significantly higher power consumption.

The researchers note that one of the main results of the work is that all the results presented were obtained directly from the equipment. In much of the earlier work on memory computing chips, AI test results were often derived in part from software simulations.

Next steps include improving the architecture and circuitry and scaling the design for nodes with more advanced technologies. Engineers also plan to look at other applications, such as impulse neural networks.

Note: Published in open access in the journal Nature.