Posits4Torch: Field-Programmable Gate Array-Enabled Posit Quantization Customization for PyTorch—A CenterFusion Case Study

The demand for energy-efficient deep learning models for autonomous driving highlights the potential of alternative number systems such as Posits. Posits are designed to improve precision, space efficiency, and dynamic range over traditional floating-point representations. However, the adoption of Posits is hindered by a lack of implementation tools. Therefore, we introduce Posits4Torch, a toolchain that integrates Posit quantization into PyTorch, enabling hardware acceleration of Posit arithmetic on field-programmable gate arrays (FPGAs). We then present a case study applying Posits4Torch to CenterFusion, a 3D object detection model benchmarked on the nuScenes dataset. Our results show the FPGA implementation achieved a 915× speedup over CPU-emulated Posit arithmetic for an 8-bit Posit-quantized CenterFusion model, though host-to-FPGA data transport overheads make it 109× slower than GPU floating-point execution. Moreover, a degradation of up to 9.8% in Mean Average Precision (mAP) and 7.7% in the nuScenes Detection Score (NDS) was observed. For 8-bit integer quantization, this degradation was up to 13.5% in mAP and 10.7% in NDS. These results underscore the trade-offs in low-precision computation. Finally, our work bridges the gap between software deep learning frameworks and hardware acceleration for Posit-based models. Future work aims to improve the hardware acceleration pipeline.