You are here : Home > Denser, faster memory for artificial intelligence


Denser, faster memory for artificial intelligence

​Advances in memory—increasing the amount of information that can be saved for a given surface area and integrating memory closer to computing resources—could bring benefits to AI. CEA-Leti, a CEA Tech institute, recently presented two major innovations at the IEDM 2020* conference.

Published on 18 December 2020

​The world of computing memory is complicated and hierarchical. Resistive memory, or RRAM, offers a good tradeoff between density and speed. This type of memory also has the advantage of being easy to integrate on logic circuits. These characteristics make RRAM particularly appealing for embedded scenarios. CEA-Leti's years of experience with advanced memory technologies recently led to several major innovations, which the institute presented at IEDM 2020*.

Researchers from CEA-Leti presented two papers on their latest advances in RRAM and the two approaches that enabled them to obtain never-before-seen densities. First, they used monolithic 3D integration to stack the different memory elements and achieve very high interconnect densities. They also made improvements to the materials used and optimized memory programming to create intermediary resistance states—instead of two states, they created a gradient of nine states per memory element. This boosted the amount of information that can be stored by each memory element threefold. The memory can also be used in analog, rather than digital, mode, which makes it suitable for neural networks, for example.

This new memory technology also enables very tight integration with logic circuits. The researchers also demonstrated that it is even possible to perform a few simple operations inside the memory itself. This could increase processing speeds by a factor of 20 and slash energy consumption by a factor of 40. These innovations could also support more complex computing in embedded objects and create new opportunities for AI.

*A leading annual conference on electronic devices

Top page