NIFA: Nonlinear IMC enhanced FPGA for efficient ML inference

arXiv cs.AI · 6 hours ago

Researchers propose NIFA, an FPGA architecture with a ReRAM-based in-memory computing block that uses analog content-addressable memories instead of ADCs to perform nonlinear operations and dynamic matrix multiplication directly in hardware. The design achieves 40x higher energy efficiency on CNNs and 1.9x on Transformers compared to conventional approaches. This extends in-memory computing benefits to Transformer models that require frequent nonlinear operations, previously limited to static-weight models.