LogicNets: Co-Designed Neural Networks and Circuits for Extreme-Throughput Applications

Deployment of deep neural networks for applications that require very high\nthroughput or extremely low latency is a severe computational challenge,\nfurther exacerbated by inefficiencies in mapping the computation to hardware.\nWe present a novel method for designing neural network topologies that directly\nmap to a highly efficient FPGA implementation. By exploiting the equivalence of\nartificial neurons with quantized inputs/outputs and truth tables, we can train\nquantized neural networks that can be directly converted to a netlist of truth\ntables, and subsequently deployed as a highly pipelinable, massively parallel\nFPGA circuit. However, the neural network topology requires careful\nconsideration since the hardware cost of truth tables grows exponentially with\nneuron fan-in. To obtain smaller networks where the whole netlist can be\nplaced-and-routed onto a single FPGA, we derive a fan-in driven hardware cost\nmodel to guide topology design, and combine high sparsity with low-bit\nactivation quantization to limit the neuron fan-in. We evaluate our approach on\ntwo tasks with very high intrinsic throughput requirements in high-energy\nphysics and network intrusion detection. We show that the combination of\nsparsity and low-bit activation quantization results in high-speed circuits\nwith small logic depth and low LUT cost, demonstrating competitive accuracy\nwith less than 15 ns of inference latency and throughput in the hundreds of\nmillions of inferences per second.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC