Technical Papers Archive

MakoXC: Rearchitecting DFT Exchange–Correlation with Matrix-Aligned and Knowledge-Organized Sparsity


Authors: Haozhi Han (School of Computer Science, Peking University); Fusong Ju and Jing Bai (Microsoft Research); Xiang Zhao (College of Computer Science and Technology, Ocean University of China); Liang Yuan and Yunquan Zhang (Chinese Academy of Sciences); Ting Cao and Yunxin Liu (Institute for AI Industry Research, Tsinghua University); Yifeng Chen (School of Computer Science, Peking University); and Kun Li (Institute for AI Industry Research, Tsinghua University)

Abstract: Density functional theory (DFT) is essential to materials science and drug discovery, yet exchange–correlation (XC) evaluation remains a major bottleneck. Linear-scaling methods exploit electronic nearsightedness but introduce irregular sparse workloads that underutilize modern AI accelerators. We present MakoXC, a matrix-aligned XC engine that restructures nearsightedness-induced sparsity into regular accelerator-friendly computation. MakoXC combines matrix-aligned cells to organize interactions into dense clusters, sparsity-guided activation to expose implicit sparsity for practical linear scaling, and a kernel-fused pipeline to consolidate fragmented workloads into a compute-intensive execution path. MakoXC achieves average speedups of 67.8x over standard XC evaluation and 4.7x over state-of-the-art linear-scaling methods. Integrated into a production-grade commercial DFT package, MakoXC scales to ubiquitin (1,231 atoms, def2-SVP) on 64 GPUs, completing end-to-end DFT in under five minutes. MakoXC demonstrates how scientific workloads can achieve genuine low complexity while maximizing parallel efficiency on AI accelerators.


Presentation: file

Back to Technical Papers Archive Listing