A domestic research team has developed a system capable of accelerating artificial intelligence (AI) learning and inference at approximately 5,000 trillion calculations per second, which is expected to be utilized in large-scale AI services such as autonomous driving, cloud computing, and data centers.

The Electronics and Telecommunications Research Institute (ETRI) announced on the 2nd that it has developed an artificial intelligence system capable of performing approximately 5,000 trillion calculations per second with low power consumption using high-performance AI semiconductor chips.

Following the unveiling of the NPU-based AI semiconductor chip, AB9 (Aldebaran 9), last year, ETRI successfully developed an AB9-based board and artificial intelligence system this year.

It is intended to be fully utilized in high-performance servers that provide AI application services such as autonomous vehicles, cloud, data centers, and people, objects, and voice recognition.

The NPU board 'ABrain-S' developed by the research team is based on AB9 and has a unique design that is small in size and consumes low power.

Artificial intelligence algorithmFor Lee, a memory capable of storing input/output data of up to 16 GB and an interface for fast data transfer were also applied.

The research team reported that maximizing the integration of NPU boards is key to implementing high-performance, high-efficiency servers.

Up to 20 NPU boards with AB9 built-in can be installed in one server node.

Compared to existing systems, space and power efficiency have been improved while the price has also been lowered.

This is because AB9 has a performance of 40 trillion operations per second in a small area the size of a coin, while consuming very little power, at around 15W.

Based on this, ETRI created ArtBrain-K, an artificial intelligence system that consists of eight server nodes stacked in a rack server format.

The developed system delivers a performance of up to 5 petaFLOPS.

This means that each server can perform approximately 5,000 trillion calculations per second.

It has approximately four times the computational performance and seven times the power efficiency compared to existing GPU-based AI servers.

When ArtBrain is applied to data centers, processing capacity and speed will be greatly improved.

Therefore, it is expected to be utilized in areas that require enormous computing resources for data processing and learning, such as the Transformer series artificial intelligence algorithm and the Huge Neural Network.

In addition, ETRI has released the SW development environment tool 'AIwareRT' on Github to facilitate the development of AI algorithms.

It provides the basic structures, algorithms, simulators, optimization tools, etc. required for programming in the form of a library so that even those unfamiliar with programming languages can easily use it.

Han Jin-ho, head of ETRI's AI Processor Research Lab, said, "We possess fundamental technology in the non-memory semiconductor field through our own AI semiconductor development. We have successfully enhanced the competitiveness of our technology by developing an NPU board equipped with AI semiconductors, an NPU server system, and related software."

Currently, this technology is being transferred to semiconductor companies and AI hardware companies, and is being applied to airport automatic immigration systems for facial recognition and immigration security.