With Peking University, Microsoft Research and the Chinese Academy of Sciences, AIR’s AIoT team proposes MakoXC, which reorganises the irregular sparsity that electron nearsightedness creates into regular, matrix-aligned GPU operations. Its exchange-correlation module is on average 67.8× faster than a dense baseline and 4.7× faster than state-of-the-art linear-scaling methods; the full DFT system runs 3.94× faster and computes the energy of a 1,231-atom ubiquitin system in under five minutes on 64 A100 GPUs. The paper was accepted to SC’26.
https://mp.weixin.qq.com/s/JwN5XQppjFvLexsnEe-UhA