Vietnamese scientists make AI chips simulate human brain mechanism

The EDABK-Bain chip by the development team won the first prize in the 3rd Silicon Design Challenge chip contest, organized by Efabless and Google co-funded in 2024.

Assoc. Prof. Nguyen Duc Minh, Leader of the research team, said that the core idea is to simulate the structure of the brain to design the new AI chip generation, in order to minimize energy consumption. Modern AI systems like Openai’s GPT need tens of thousands of GPUs to train, consuming huge amounts of electricity. Meanwhile, the human brain only consumes about 20W and still processes complex information effectively. “We want to learn by this mechanism to optimize the performance of AI chip,” said Assoc. Prof. Minh said.

The traditional chip architecture is based on the Von Neumann model, with the separation between memory and processor. However, the research team of the Polytechnic University applied the architecture “In-Memory Computing” (calculated at the memory), helping the calculation element and memory are close together. This architecture simulates the pulse neuron network, similar to the activity of the biological nervous system, helping to speed up processing and minimize energy consumption.

Besides the application of new architecture, the group also takes advantage of artificial intelligence to support chip design. Dr. Hoang Phuong Chi, a member of the research team, said: “The group uses AI tools, including chatgting, to optimize the design process, helping to shorten the development time from the idea to the finished product in just 12 days.” However, Dr. Chi emphasized that to achieve this achievement, the group had more than three years of foundation, starting from 2020 and published the first building at the European Science Conference by 2021.

The EDABK-Bain chip is smaller by the development group than IBM or Intel’s commercial products. It is built on 130 nm process with 256 neurons and 65,000 neurological joints, while the most advanced chips now reach millions of neurons. Although still limited in size, but according to Assoc.

Currently, large corporations like IBM and Intel are also pursuing the direction of developing AI chips inspired by the brain. However, the group of Hanoi University of Technology focuses on building an open-source open system, including the design, synthesis and model training tools. This is a factor that helps the group win awards at the international chip design contest.

Assoc.Prof.Dr. Nguyen Duc Minh said that the chip can be integrated into smart wearing devices such as enhanced actual glass (AR), Electrocardiography, and Electrical Electrical Measures to detect early biological abnormalities. Thanks to low energy consumption, the chip can operate for a long time without a large supply.

Currently, the group continues to research and develop more advanced AI chip versions, contributing to the global trend in the construction of energy -saving systems and capable of handling complex AI tasks at the edge (Edge AI) – that is, processing information right at the device without a central server, helping to respond faster but low energy consumption.