A new breakthrough memory technology that could boost AI is about to go mainstream: rivals meet at major event to discuss IGZO DRAM as in-memory computing computers that move one step closer to reality
Researchers are experimenting with a number of different forms of memory that may be better suited to AI, with indium gallium zinc oxide (IGZO) being the latest to gain attention.
IGZO-based 2-transistor 1-capacitor (2T1C) technology is typically found in displays, but research organization imec has identified its potential for Analogue In-Memory Computing (AIMC).
The AIMC approach addresses the limitations of traditional digital computing, especially speed and energy efficiency, by performing in-memory computing tasks itself using analog technology. This minimizes power consumption and accelerates calculation speed.
Denser memory array
The main advantage lies in the parallel processing and storage of data in analog format in memory, which provides a faster, more efficient and energy-saving way of computing. Essentially, memory itself becomes part of the computing process, negating the need for data transfer between separate units.
IGZO DRAM cells hold promise for analog in-memory computing due to their significantly lower standby power consumption. Additionally, IGZO transistors can be incorporated into the back-end-of-line (BEOL) of the chip, allowing placement on top of the peripheral circuitry in the front-end-of-line (FEOL). This results in a denser memory array without a FEOL footprint.
At the recent 2023 International Memory Workshop (IMW), imec teams addressed some of the remaining challenges, strategies for optimizing gain cell retention time, and demonstrations of successful MAC operation in an array configuration.
It wasn’t the only company discussing the technology either, as Samsung also shared its research there.
You can read more about the subject on the website imec site, but the researchers conclude that the IGZO-based 2T1C and 2T0C amplification cells (a capacitor-less variant) exhibit exceptional properties for AIMC. Compared to traditional SRAM-based technology, they provide superior energy efficiency and compute density for machine learning applications, especially during the inference phase. The 2T0C cells excel even further in surface efficiency.