TRINE: A Token-Aware, Runtime-Adaptive FPGA Inference Engine for Multimodal AI

In three linesTRINE is an FPGA accelerator and compiler for end-to-end multimodal inference (ViT, CNN, GNN, transformers) without reconfiguration. It unifies layers as matrix operations, switches between systolic and SIMD architectures at runtime, and applies in-stream token pruning. On Alveo U50 and ZCU104, it achieves 22.57x latency reduction vs RTX 4090 while consuming 20-21 W.

Summary generated by Claude — human-verified