Breaking the Memory Wall: Nanoelectromechanical Switches for the Next Generation of AI SoCs

Researchers are tackling the 'memory wall' in AI SoCs by developing nanoelectromechanical switches for eDRAM. These innovations aim to reduce power consumption and increase bandwidth for the next generation of high-performance chips.

Share
Breaking the Memory Wall: Nanoelectromechanical Switches for the Next Generation of AI SoCs

As AI models grow in complexity, the "memory wall"—the performance gap between the processor and memory—has become the primary bottleneck in semiconductor design. Researchers from UC Berkeley, Samsung, and Intel are now exploring nanoelectromechanical (NEM) switches to create more efficient embedded DRAM (eDRAM). This approach aims to provide the density of DRAM with the speed and low power consumption required for modern AI and High-Performance Computing (HPC) SoCs.

Traditional SRAM (Static Random Access Memory) takes up significant die area, while standard DRAM often lacks the necessary bandwidth for real-time AI processing. The NEM-gated gain cell eDRAM offers a potential solution by significantly reducing leakage current and improving energy efficiency. This is particularly vital for "Edge AI" applications where power constraints are as important as raw throughput.

Furthermore, as the industry moves toward RibbonFET and PowerVia architectures, the integration of memory IP becomes even more complex. The focus is shifting toward 3D-IC packages and heterogeneous integration to keep data closer to the compute cores. By breaking the memory bottleneck at the transistor level, these new switches could enable a new class of AI chips capable of processing massive datasets with a fraction of the current energy footprint.


Source: Semiconductor Engineering