Terafly : A Multi-Node FPGA Based Accelerator Design for Efficient Cooperative Inference in LLMs

In this paper, we propose Terafly, a multi-node accelerator design tailored for efficient Large Language Model (LLM) deployment and inference. Conventional accelerator architectures struggle to effectively handle both the prefill and decode stages during inference. To address this limitation, we introduce a hybrid spatial-temporal architecture that combines the high-throughput advantages of spatial architectures with the flexibility of temporal architectures, enabling it to accommodate the diverse inference patterns of LLMs. In addition, we propose a generation framework to streamline the customization of our LLM-friendly design for various deployment scenarios. Within this framework, users can specify their requirements such as model type, target platform, and performance goals. The framework then generates multiple accelerator nodes and maps them to distinct Super Logic Regions (SLRs) within a single FPGA, enabling cooperative inference under a model parallelism scheme. Through experiments, our generated accelerator can be easily deployed on both Alveo U250 and U50lv cards, serving models ranging from OPT-350M to OPT-1.3B under various performance settings. Notably, when running OPT-1.3B using the generated dual-node accelerator on a single Alveo U50lv card, we achieve an average 1.1x speed-up and a 3.4x improvement in energy efficiency compared to the Nvidia A100 GPU.