It is expected that the development of Korea's first analog DRAM-PIM-based AI semiconductor by a domestic research team will contribute to preempting the growing AI model market such as ChatGPT.

The Ministry of Science and ICT announced that Professor Yoo Hoi-jun's research team at the Korea Advanced Institute of Science and Technology (KAIST) has developed 'DynaPlasia*,' a PIM semiconductor that performs AI calculations by directly integrating a calculation unit inside a DRAM memory cell for the first time in Korea.

PIM (Processing-in-Memory) is a technology that performs data processing by integrating memory and processor within a single chip. It has emerged as a hidden card for next-generation AI semiconductors as it can solve the problems of data bottlenecks and excessive power consumption that occur in existing computing structures (Von Neumann Architecture) where memory and processors are separated.

At a briefing on research results related to PIM semiconductors on the 14th, KAIST ICT Professor Yoo Hoe-joon said, "TSMC has been leading the way so far, but we have been researching Samsung and SK Hynix with the idea that our country, a memory powerhouse, should take the lead and focus on memory."

He also said about this research result, "However, previous PIMs were not the ultimate PIM, but took the form of PNM, that is, Process-Near-Memory rather than Process-In-Memory, and what was announced this time is a true PIM, a structure in which the processor is within the memory."

In particular, he said, "It has memory functions as well as a processor, and it is an analog computing method, and it is called 'triple-mode' because one cell can be a memory, a processor, and a capacitor for an ADC."

According to Professor Yoo, although PIM semiconductors have been developed in the past, most of them were either SRAM-PIM methods that require more than eight transistors per cell, or, even if they were implemented as DRAM-based PIMs like existing PIMs, they were digital PIMs (Near Memory PIMs) that placed the calculation units close to the outside, rather than inside, of the memory cell array.

This digital PIM method reduces data bottlenecks by reducing the distance between memory and the processor and increasing bandwidth, but it does not increase computational performance by integrating the processor directly inside the memory cell.

This time, the research team announced 'Dynaplagia (D)ynaPlasia) has optimized the entire hardware structure. It is an analog DRAM-PIM-based AI semiconductor, and the cell consists of only three transistors. It integrates the calculation unit inside the memory cell and improves the integration and calculation function by utilizing an analog calculation method. All memory cells can operate in parallel through tolerance computing that eliminates leakage current in the internal multiplication logic.

Ultimately, it is explained that it shows 15 times higher data processing capacity with approximately 300 times higher parallelism compared to the existing digital DRAM-PIM method.

In addition, in existing analog PIM semiconductors, memory, arithmetic unit, and data converter were implemented separately and a fixed hardware structure was used, but in this study, a 'triple-mode cell' was developed in which a single cell can support the functions of memory, arithmetic unit, and data converter simultaneously.

With the dynamic core formation architecture that forms the hardware structure according to the actual AI operation through the 'triple-mode cell', an efficiency approximately 2.5 times higher than that of existing analog PIM semiconductors can be achieved.

Meanwhile, this research was conducted at the 'PIM Semiconductor Design Research Center (hereinafter referred to as 'PIM-HUB')' established through the 'PIM Artificial Intelligence Semiconductor Core Technology Development (Design)' project of the Ministry of Science and ICT, and was presented at the International Solid-State Circuits Conference (ISSCC) held in San Francisco, USA last month.

'PIM-HUB', which opened in June of last year, is promoting human resource exchanges such as mutual dispatch of personnel, joint research, and joint development of educational courses for human resource development, with Samsung Electronics and SK Hynix participating as operating committee members.

KAIST Professor Yoo Hoe-Jun of the Department of Electrical and Electronic Engineering said, “This research is significant in that it not only resolves the memory bottleneck problem of existing AI semiconductors, but also develops a high-memory capacity DRAM-PIM with high processing capacity and variability. He explained the significance of the research, saying, “If full-scale commercialization is successful, it will be able to demonstrate high performance even in artificial intelligence models that have become increasingly larger and more diverse.”

The Ministry of Science and ICT’s Jeon Young-soo, the director general for information and communications industry policy, said, “PIM semiconductor technology is a field in which our country, which has strengths in memory semiconductor technology, can take the lead,” and added, “We will not spare active support for the development of not only hardware but also software and cloud-related technologies necessary to apply domestic AI semiconductors to data centers, so that we can develop world-class ultra-high-speed, low-power AI semiconductors.”