Images related to the neuromorphic chip thesis of Samsung Electronics and Harvard University. /Photo courtesy of Samsung Electronics
Researchers at Samsung Electronics and Harvard University in the United States presented a future vision for the neuromorphic chip, the next-generation artificial intelligence (AI) semiconductor technology.
Samsung Electronics announced on the 23rd local time in the UK that a paper on neuromorphic chips written by Ham Don-hee, a fellow at Samsung Electronics Advanced Institute of Technology and Harvard University professor, Park Hong-geun, Harvard University professor, Hwang Seong-woo, Samsung SDS president and Kim Ki-nam, vice chairman of Samsung Electronics was published in Nature Electronics (UK time). ) was published in
Neuromorphic semiconductor is a semiconductor that is inspired by or imitated directly from the human brain’s neural network, and aims to reproduce even higher-order functions of the brain, such as cognition and reasoning. In this paper, the electrical signals of neurons (nerve cells) in the brain neural network are measured with nano-electrodes with high sensitivity, the connection map between neurons is copied (Copy), and the unique function of the brain is analyzed by pasting the copied map into a memory semiconductor. We proposed a technological vision of a reproducible neuromorphic chip.
According to Samsung Electronics, the copy of the neural network map through ultra-sensitive measurement is made through the arrangement of nano-electrodes penetrating the neurons. By penetrating into the neuron, the measurement sensitivity is increased, and the insignificant electrical signal generated at the contact point of the neurons can be read. In this way, it is possible to map the neural network by finding those points of contact. This is a technology that Samsung Electronics has been collaborating with the Harvard University research team since 2019.
In particular, Samsung Electronics has proposed a new concept of neuromorphic semiconductor in which each memory serves as a contact point between neurons by pasting the copied neural network map into a memory semiconductor. It takes a lot of time to construct a neural network map by analyzing a huge amount of signals measured in a neural network with a computer. In this regard, a new technical point of view was also presented for quickly downloading a neural network map by directly driving the memory platform with the measurement signal.
The platform can utilize commonly used memory such as flash and another type of non-volatile memory, such as resistive memory (RRAM). Ultimately, to realize about 100 trillion neurons in the human brain as a memory network, the memory density must be maximized. To this end, we proposed the use of cutting-edge semiconductor technologies such as 3D flash stacking technology and 3D packaging through TSV (through silicon electrode) applied to high-performance DRAM.
Samsung Electronics explained that this study is also meaningful in that it showed a vision for the next-generation AI semiconductor by combining neuroscience and memory technology with the participation of technology leaders from academia and industry. It plans to secure technological leadership in the next-generation AI semiconductor field by continuously focusing on neuromorphic research based on its existing semiconductor technology capabilities.
“The bold approach proposed in this paper will help to broaden the boundaries of memory and system semiconductor technology and further develop neuromorphic technology,” said Donhee Ham, a fellow at Samsung Electronics Advanced Institute of Technology.
Copyright owner ⓒ ‘Economic News of People Dreaming of Success’ Money S, unauthorized reproduction and redistribution prohibited