Search papers, labs, and topics across Lattice.
This paper introduces a method for mapping Differentiable Logic Gate Networks (DLGNs) directly to CMOS standard cell libraries, enabling custom silicon implementations of neural networks. A novel loss function is proposed to optimize the DLGN for minimal circuit area, indirectly reducing power consumption. The authors demonstrate a silicon implementation of a DLGN in the SkyWater 130nm process, achieving 97% accuracy on MNIST classification at 41.8 million inferences per second with 83.88 mW power consumption.
Neural networks made of logic gates can now be directly compiled to silicon, achieving impressive MNIST classification speeds with low power consumption.
Recent work in the machine learning literature has demonstrated that deep learning can train neural networks made of discrete logic gate functions to perform simple image classification tasks at very high speeds on CPU, GPU and FPGA platforms. By virtue of being formed by discrete logic gates, these Differentiable Logic Gate Networks (DLGNs) lend themselves naturally to implementation in custom silicon - in this work we present a method to map DLGNs in a one-to-one fashion to a digital CMOS standard cell library by converting the trained model to a gate-level netlist. We also propose a novel loss function whereby the DLGN can optimize the area, and indirectly power consumption, of the resulting circuit by minimizing the expected area per neuron based on the area of the standard cells in the target standard cell library. Finally, we also show for the first time an implementation of a DLGN as a silicon circuit in simulation, performing layout of a DLGN in the SkyWater 130nm process as a custom hard macro using a Cadence standard cell library and performing post-layout power analysis. We find that our custom macro can perform classification on MNIST with 97% accuracy 41.8 million times a second at a power consumption of 83.88 mW.
