Performance Evaluation of AMD Milan‑X 7773X CPU for HPC Workloads (WRF and OpenFOAM) on NF5468A5 Server

The AMD Milan‑X 7773X, released in 2023, is an upgraded third‑generation EPYC processor featuring 3D V‑Cache technology that triples the on‑die L3 cache to 768 MB, uses a multi‑chip module (MCM) with eight CCDs and a large I/O die, and remains compatible with LGA 4094 motherboards supporting up to 4 TB per socket.

To assess its impact on high‑performance computing (HPC) workloads, the authors selected two representative applications: the Weather Research and Forecasting model (WRF) and the computational fluid dynamics solver OpenFOAM. Tests were conducted on a浪潮 NF5468A5 4U GPU server equipped with two AMD CPUs (Rome 7742, Milan 7543, Milan‑X 7773X), eight NVIDIA A100/A30/A40 GPUs, 32 × 64 GB DDR4‑3200 memory (up to 8 TB total), and a PCI‑e 4.0 interconnect.

Platform specifications

Software environment

Software version

Operating System

RedHat Enterprise Linux 8.3.2011 x86_64

Compiler

Intel Compiler 2021.2.0

Parallel environment

Intel MPI 2021.2.0

Application software

WRF‑v3.9.1, OpenFOAM‑v1906

WRF benchmark

The WRF test used a two‑level nested grid (12 km and 4 km resolution) with 425 × 300 × 35 and 1150 × 802 × 35 points respectively, running for a 3‑hour simulated period. Compared with the Rome 7742 baseline, Milan 7543 achieved a 1.14‑1.27× speed‑up, while Milan‑X 7773X delivered a 1.23‑1.34× improvement. L3 cache miss rates dropped from 50‑70 % on Milan 7543 to 25‑55 % on Milan‑X.

OpenFOAM benchmark

The OpenFOAM motorBike case (simpleFoam solver, SST‑k‑ω turbulence model) generated a mesh of ~10.3 million cells. Using the same core counts, Milan 7543 provided a 1.23‑1.28× speed‑up over Rome, while Milan‑X achieved 1.34‑1.80× acceleration. L3 cache miss rates fell from ~40 % on Milan to 20‑30 % on Milan‑X.

Conclusion

The experiments confirm that the ultra‑large L3 cache of AMD Milan‑X CPUs significantly alleviates memory‑bound bottlenecks in HPC workloads, delivering up to 34 % faster WRF runs and up to 80 % faster OpenFOAM simulations compared with the previous generation Rome platform.

Future work will explore multi‑node scaling of Milan‑X in larger clusters.