Presentation
Atto-Spinformer: An Energy-Efficient Magneto Electric Spin Orbit Logic-Based Compute-in-Memory Transformer Architecture
DescriptionTransformers form the foundation of modern natural language processing, but their performance is limited by GPU memory bandwidth and high energy consumption. In-memory computing (IMC) architectures mitigate data transfer overhead, yet remain constrained by power-hungry interfaces. We present a novel ultra-low-power Magneto-Electric Spin Orbit (MESO)-based IMC architecture optimized for dot product operation acceleration within transformers. By exploiting device non-volatility and a custom transmission-gate-based pulse application scheme, our design eliminates DACs at the CMOS-MESO interface and reduce interface power by 20.3x. We propose a custom computation scheme that significantly lowers the floating-point operations required by IMC architectures to compute outputs. Combined with the low-power MESO devices, these techniques enable our accelerator to boost computational efficiency by 4.4x relative to conventional IMC designs and 2.6x relative to the NVIDIA A100 GPU.
Event Type
Work in Progress
TimeMonday, July 276:53pm - 6:53pm PDT
LocationExhibit Hall