Composed of 576 neurons, launching in July with 3 boards
Artificial Intelligence Requires a Hardware-Based, Not Software-Based Approach

In the age of artificial intelligence, the key is the chip. There is no longer a place for software-based research.

At the 'AI New Industry Convergence Forum' hosted by the Korea Institute for Industrial Technology Promotion, Nepes Chairman Lee Byung-koo stated that artificial intelligence has become closely integrated into our lives and made the following remarks regarding methods for its popularization. Following this, Nepes Executive Director Ahn Jung-ho introduced the 'NM500,' an AI chip set to be released next month, noting that "there are many advantages to approaching AI from a hardware perspective."

Nepes, a semiconductor packaging company, designed a system semiconductor chip that implements artificial intelligence by signing a neuromorphic architecture licensing agreement with the U.S. semiconductor technology company General Vision (GV).

"Neuro" means nerve and "morphic" means shape; thus, neuromorphic is a chip that shapes nerves, or brain cells. The human brain is a hardware configuration in which individual neurons make judgments and are connected by synapses to process information simultaneously in parallel. The NM500 is a hardware method in which a single chip consists of 576 neurons, and after dividing the area to perform the desired training, the data is combined to derive a result.

Neuromorphic has eight characteristics, including the ability to simultaneously transmit input data to all neurons, constant response speed even as the number of neurons increases, exclusion of neurons showing low response from competition, automatic alignment of results based on the degree of neuron response, dynamic addition of new knowledge, spontaneous inhibition of incorrect learning, software implementations are not neuromorphic, and fast and free uploading and downloading.

Executive Director Ahn stated, “The chip scheduled for release consists of 576 neurons. With 576 neurons, it is capable of powering most wearable devices. It operates on low energy of 0.3W, and the chip size is very small at 4.5mm,” adding, “The small size of the chip means that it can intelligentize electronic devices requiring intelligence and all hardware.”

The NM500 chip can utilize all data that a computer can accept, such as images, videos, text, and signals. Once the data is vectorized and stored in the chip, the chip itself determines the subsequent results. These results can then be used for verification, classification, motion tracking, and more.

Also, if more neurons are needed to analyze more data, it can be expanded by inserting other chips next to it or stacking them on top. The number of neurons can be expanded to over a million, and the response speed remains constant regardless of the number.

When combined with sensors, it is expected to be utilized in various applications, including not only electronic devices but also semiconductor inspection processes, X-ray image interpretation, automotive electronics, and home cognitive sensors.

We have also developed our own software, “Knowledge Builder,” which is a tool that trains intelligence according to field conditions and enables immediate use. The application boards scheduled for release are basically three types: Neuroshield, which allows for quick and easy experience with neurons; Neurobrick, which allows you to back up the number of neurons as much as you want; and Orange Brain Card, which has a built-in image sensor and is suitable for full-scale development.

Executive Director Ahn expressed his expectations, stating, “We plan to supply the NeuroShield product as open source. You are free to modify it to suit your needs and sell it. The reason for this openness is that we are currently lagging behind in the field of artificial intelligence software. However, since hardware is one of our strengths, we hope to become a global leader.”