• Memory wall hampers AI; data movement between processor and memory costly. • Resistive RAM (RRAM) offers in‑memory computing but traditional types unstable. • UCSD researchers redesigned RRAM switching mechanism to enable reliable analog matrix operations. • New RRAM can perform multiply‑accumulate via current flow, eliminating high‑voltage filament formation. • Eliminates noisy filament process, removes selector transistor, enabling 3D stacking and CMOS integration. • Demonstrated running a learning algorithm on the new RRAM, promising faster, energy‑efficient neural networks.
Article Summaries:
- UCSD researchers have demonstrated a new bulk‑mode resistive RAM (RRAM) that could help overcome the memory wall limiting AI workloads. By eliminating the noisy filament‑forming step and the need for selector transistors, the team created 40‑nanometer cells that switch an entire layer between high and low resistance. An eight‑layer stack can assume 64 distinct resistance levels, with each cell operating in the megaohm range-well suited for parallel matrix operations in neural networks. The researchers assembled a 1‑kilobyte array without selectors and successfully ran a learning algorithm on it, suggesting bulk RRAM may enable more efficient on‑chip computation for machine‑learning tasks.
- UCSD researchers presented at IEDM a new bulk‑type resistive RAM (RRAM) that could help overcome AI’s memory‑wall bottleneck. Unlike conventional filament‑based RRAM, the devices switch an entire layer between high‑ and low‑resistance states, eliminating the high‑voltage filament‑forming step and the need for selector transistors. The team shrank the cells to 40 nm, stacked up to eight layers, and achieved 64 distinct resistance levels per cell-well above the kilo‑ohm range of traditional RRAM. A 1‑kB array of these cells, requiring no selectors, was shown to run a learning algorithm, demonstrating the potential for in‑memory matrix operations that could accelerate neural‑network workloads.
Sources: