Answers are grounded in this paper and cite their sources.
Point Cloud Neural Network (PCNN) plays an essential role in various 3D applications, with some of them even being time-sensitive and safety-critical. However, the large scale of unordered points with lengthy features results in heavy computational workloads, making them far from real-time processing. To address this challenge, we propose PQuant, a Morton-code-based fine-grained quantization for accelerating both point-based and voxel-based networks. Specifically, we utilize Morton-code to capture the inherent point similarity and the voxel distribution. Considering the similarity among near points’ features, we decompose features into base and offsets, where the offsets typically exhibit a narrow dynamic range and can thus be quantized using low precision. For the voxel-based networks, spatial location indicated by Morton-codes allows PQuant to apply high precision representation to distant voxels, which are generally sparser and more sensitive to quantization. From the hardware aspect, PQuant integrates a reconfigurable mixed-precision systolic array designed to handle both dense and sparse mixed-precision convolutions in PCNNs. Extensive experiments demonstrate that PQuant achieves 3.7 × and 3.5 × speedup and 5.2 × and 6.0 × energy savings over state-of-the-art PCNN accelerator PointAcc for the point-based and voxel-based networks, respectively.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: