Through its exceptional capabilities, ChatGPT is transforming numerous sectors beyond just education. With its advanced AI language model, this platform offers a dynamic question-and-answer-based interaction to perform different tasks such as translation, coding, paper writing, and much more.
To minimize mistakes, the AI system requires a lot of training, which necessitates frequent transfers of data between processors and memory. On the other hand, traditional computer systems separate the storage and computation of information using the von Neumann architecture, resulting in heightened energy consumption and delays in AI computations. To solve this issue, researchers have come up with semiconductor technologies that are well-suited for AI tasks.
“The significance of my research team’s accomplishment is that we were able to overcome the constraints of traditional AI semiconductor technologies that were purely focused on material development. We used materials that were already in mass manufacturing to do this.”
Professor Chung
The utilization of indium gallium zinc oxide (IGZO), a prevalent oxide semiconductor found in OLED displays, has led to the creation of an advanced AI semiconductor apparatus. Professor Yoonyoung Chung, Professor Seyoung Kim, and Ph.D. candidate Seongmin Park from the Department of Electrical Engineering, Department of Semiconductor Engineering, and Department of Materials Science and Engineering at POSTECH have spearheaded the research team responsible for this development.
The device’s performance and power efficiency have been outstanding, as confirmed by the latest research.
Oxide semiconductors are utilized to operate the AI synaptic device. By charging or discharging the storage node through the write transistor, the read transistor’s conductance can be manipulated. After the weight updates are finished, the programmed weights are retained when the writing transistor is deactivated. Recognition: POSTECH.
For AI operations such as ChatGPT to function optimally, calculations must take place in the same memory that stores information. However, past chip technology for AI wasn’t up to snuff, lacking key features like linear and symmetrical programming as well as consistency in order to improve accuracy.
In their pursuit of a material for mass-produced AI computations with consistency, longevity, and high accuracy, the research team identified IGZO as a prime candidate. IGZO, which is composed of four atoms in predetermined ratios of indium, gallium, zinc, and oxygen, possesses exceptional electron mobility and leakage current properties. These qualities have secured its position as the backbone of OLED displays.
The researchers built a unique synapse device with two interconnected transistors and a storage node. The ability to precisely control the charging and discharging speed of the storage node led to the creation of an AI semiconductor capable of meeting high-level performance standards.
The proposed synthetic synapse devices showcase characteristics of potency and depression that are influenced by the input voltage. They are based on an oxide semiconductor and emulate the synaptic functions of the human brain. These devices have exceptional linearity and symmetry in terms of their potentiation and depression, and their output current is directly proportional to the number of input pulses received. The output current can be used to perform artificial intelligence (AI) tasks with remarkable accuracy and minimal power consumption owing to its near-proportional correlation with the input voltage. POSTECH deserves credit for its efforts in this regard.
The optimization of synaptic devices is a critical aspect when it comes to integrating them into a large-scale AI system. To achieve this, the researchers discovered that ultra-thin film insulators used in transistors can regulate current output, making them ideal for this purpose.
The potential for high-accuracy AI systems in the future was proven by the researchers’ success in training and classifying handwritten data using the recently developed synaptic device, which resulted in an impressive accuracy rate of more than 98%.
The IGZO 2T synaptic devices were put through a simulation test for MNIST classification and produced promising results. By utilizing oxide semiconductors, these AI synaptic devices demonstrated excellent linear and symmetric weight programming abilities, promoting precise training procedures that minimize potential artificial intelligence-related setbacks. In a demonstration for the recognition of random handwriting, the devices were able to achieve an inference accuracy of nearly 98%, a noteworthy accomplishment that is comparable to an ideal mathematical calculation’s accuracy. The credit goes to POSTECH.
We managed to overcome the restrictions of conventional AI semiconductors, which were primarily focused on material advancement, stated Professor Chung, discussing his team’s achievement. Their approach included the use of readily available materials in mass production. Additionally, they were able to develop linear and symmetrical programming attributes through a new structure that utilized two transistors as a single synaptic device.
The potential to enhance the precision and effectiveness of AI has been demonstrated through our triumphant implementation and advancement of innovative AI semiconductor technology.
Advanced Electronic Materials ran an article on the research.
More information: Seongmin Park et al, Highly Linear and Symmetric Analog Neuromorphic Synapse Based on Metal Oxide Semiconductor Transistors with Self‐Assembled Monolayer for High‐Precision Neural Network Computation, Advanced Electronic Materials (2022). DOI: 10.1002/aelm.202200554
