Researchers at Peking University have ported the ABACUS materials simulation package to Huawei Kunpeng 920 processors, achieving speedups of up to 3.9 times over a mainstream x86 CPU at 28 cores. The team tuned 19 core operators within the software to run on the Arm-based architecture, recording individual component accelerations reaching 10.7 times.
Key points
- A Peking University research group adapted the open-source ABACUS density functional theory software for Huawei’s Kunpeng 920 server processor.
- The researchers optimised 19 core computational operators, using the processor’s vector and matrix hardware units.
- Full simulation workloads ran between 1.3 and 3.9 times faster on the Kunpeng 920 than on a mainstream 28-core x86 central processing unit.
- The port forms part of a wider effort by Chinese institutions to run critical scientific computing workloads on domestic silicon.
Optimising ABACUS for Arm architecture
ABACUS, short for Atomic-orbital Based Ab-initio Computation at UStc, is an open-source density functional theory software package used to calculate the electronic structures of molecules and materials. The code is widely used by researchers to predict physical properties, design catalysts, and model crystal structures without relying on physical lab experiments.
Scientific modelling applications of this type consume massive amounts of floating-point computing power. Traditionally, research laboratories run these calculations on x86 server chips from Intel or AMD. Porting ABACUS to the Huawei Kunpeng 920 required adjusting the lower-level numerical routines to target the processor’s custom Arm architecture.
The Peking University team focused on 19 computational operators that handle the bulk of the matrix calculations and numerical integrations in density functional theory. By tailoring these routines to Kunpeng’s vector hardware and matrix units, the researchers eliminated execution bottlenecks that typically slow down non-native code.
Benchmark results and performance gains
In side-by-side benchmark tests across standard materials simulation tasks, the tuned software on the Kunpeng 920 outperformed a 28-core x86 system. At identical core counts of 28 cores, the overall calculation runs showed speedups ranging from 1.3 times to 3.9 times.
Specific mathematical operators saw even higher gains. The team reported that certain individual routines within the 19 tuned operators ran up to 10.7 times faster after being rewritten to use the chip’s internal matrix and vector extensions. Matrix multiplication and tensor contractions are fundamental to density functional theory, making specialized execution units particularly effective for these workloads.
Density functional theory calculations scale heavily with the number of atoms in a simulation. The reported efficiency improvements allow researchers to either model larger atomic systems within the same compute budget or run standard calculations in a fraction of the time.
China’s push for domestic scientific software
The successful port of ABACUS highlights an ongoing drive within Chinese universities and supercomputing centres to build an independent software and hardware ecosystem. US trade controls have limited Chinese access to high-end chips from Nvidia and Intel, prompting institutions to invest heavily in domestic alternatives.
Huawei developed the Kunpeng 920 to serve enterprise data centres and high-performance computing clusters. While the chip has been deployed across domestic cloud providers, scientific computing packages often require years of bespoke optimisation before they match the performance of mature x86 deployments.
Scientific packages such as ABACUS, VASP, and Quantum ESPRESSO form the bedrock of modern condensed matter physics and computational chemistry. Adapting such tools to Kunpeng servers ensures that research institutions can continue materials discovery projects on locally produced silicon.
What the port means for researchers
High-performance computing clusters rely on tight integration between compilers, numerical libraries, and processor registers. By making the optimisations open-source, the Peking University team provides a blueprint for other scientific development groups adapting computational physics codes to Kunpeng processors.
The project shows that Arm-based server chips can compete effectively with conventional x86 chips in floating-point workloads when the underlying mathematical operators are systematically tuned. Materials science groups operating Kunpeng hardware can now deploy ABACUS directly without facing an emulation or architectural performance penalty.
The research team plans to continue refining the codebase and evaluating larger multi-node cluster deployments to measure how the tuned operators scale across thousands of cores.
