Beyond Memory

In today’s computing environment, higher memory bandwidth and capacity are becoming ever more important for system scaling. In the current memory-centric system era, the boundaries between compute and memory are becoming more blurred. We believe this presents an opportunity for new memory-compute options that may bring value to compute and memory simultaneously. At SK hynix, we are tackling many key challenges in order to move toward a “beyond-memory era”.

In particular, we believe that emerging memory may be important for the path beyond memory by blurring the lines between compute and memory. We are evaluating this option since ACiM cells have many commonalities with known memory cells while its unique optimization such as variation and linearity exists. We have successfully demonstrated 16 levels of RRAM-based Synapse cells platform with good set/reset characteristics which are embedded in a CMOS technology. Through SK hynix's Open Research Platform(ORP), ACiM technology opens the possibility of collaboration with various partners.

New Memory

DRAM and NAND that are currently in the market use electrical charges to perform programming and reading operations. The physics of these devices has been well-established and their commercialization required decades of research and development effort. In contrast, new memory devices perform programming and reading operations using current, where applying current changes resistance in the memory. We expect NM to bridge the gaps between DRAM and NAND in terms of performance, latency, and cost.

• Our R&D activities in the new memory area focus on using SOM(Selector-Only Memory) materials and developing integration processes to achieve better performance, reliability, cost, and scalability than those in conventional 3D cross-point memory(3DXP).

ACiM

With the recent rapid growth of large language models (LLM) such as GPT, the computational cost of training the Transformer model has expanded at unprecedented rate. In-memory computing has been of great interest for the applications of AI accelerators. Naturally NVM arrays in cross-point architectures have been extensively studied for analog-compute-in-memory(ACiM), due to its high energy efficiency via minimizing data movement.

• We are researching NVM-based ACiM cells platform, embedded in a CMOS technology and we look forward to the possibility of collaboration with various R&D partners through SK hynix's Open Research Platform(ORP).