Event

LENS accelerates robot scene memory on FPGA

1 reports · 1 sources

Haohua Que, Mingkai Liu, Tianle Zhu, and Handong Yao propose and implement LENS, which accelerates ACEsplat-style scene memory workloads on the AMD KR260 FPGA/ARM platform; the system comprises an INT16 scene coordinate regression accelerator and a fixed-point ordered 3DGS renderer. The authors state that LENS's latency, power, and quality results are all measured on-board; on-device training can reach ACEsplat-level quality, but its time and energy consumption fall short of a GPU, which the authors regard as a boundary that future training accelerators still need to break through.