FusionCIM: Accelerating LLM Inference with Fusion-Driven Computing-in-Memory Architecture
This work addresses the energy and latency bottlenecks of LLM inference for hardware accelerator designers, offering a significant improvement over existing CIM architectures.
FusionCIM proposes a compute-in-memory accelerator for LLM inference that fuses attention operations, achieving up to 3.86x energy savings and 1.98x speedup over prior CIM designs, with 29.4 TOPS/W system-level energy efficiency on LLaMA-3.
In this paper, we propose FusionCIM, an operator-fusion-driven compute-in-memory (CIM) accelerator architecture for efficient and scalable LLM inference, with three key innovations: (1) a hybrid CIM pipeline architecture that maps QKT computation on inner-product-based CIM (IP-CIM) and PV aggregation on outer-product-based CIM (OP-CIM) for efficient matrix multiplications fusion; (2) a QO-stationary dataflow that eliminates repeated KV loading in CIM and K-matrix access in buffer under transpose fusion, significantly improving data reuse on chip; and (3) a pattern-aware online-softmax mechanism that exploits distribution regularities of attention scores to reduce exponential rescaling overhead for non-linear fusion. Experimental results on LLaMA-3 model show that FusionCIM achieves up to 3.86x energy saving, and 1.98x speedup compared with prior SOTA CIM-based designs with 29.4 TOPS/W energy efficiency at the system level.