Performance Analysis and Demo of Alibaba Cloud HPC8ae Optimized Instance for CFD Simulations

Alibaba Cloud recently launched the HPC8ae optimized instance, built on the self‑developed "Feitian+CIPU" architecture and equipped with a fourth‑generation AMD EPYC Zen4 processor (3.4 GHz base, 3.75 GHz boost). The instance enhances memory bandwidth and supports eRDMA, making it well‑suited for high‑performance computing (HPC) workloads such as CFD, climate modeling, and other simulation‑intensive applications. Measured performance includes a floating‑point peak of 3.63 TFLOPS and a memory stream bandwidth of 603 GB/s, with eRDMA latency around 8 µs.

The article documents two hands‑on demos performed by Alibaba Cloud HPC engineer Cao Hang. The first demo showcases a single‑node motorcycle external‑flow simulation using OpenFOAM. The workflow involves creating an EHPC cluster, selecting the HPC8ae instance as the compute node, deploying the OpenFOAM environment, submitting the pre‑configured simulation job, and visualizing results with ParaView.

The second demo scales the workflow to a multi‑node environment, running a 60‑million‑cell wind‑farm simulation across six HPC8ae nodes. The steps mirror the single‑node process but emphasize the high‑bandwidth, low‑latency eRDMA network for efficient parallel communication. Performance highlights include a substantial single‑node speedup, one‑click deployment of large HPC clusters, and seamless front‑ and back‑end processing via the E‑HPC portal.

Overall, the HPC8ae instance demonstrates significant advantages for simulation workloads: superior single‑node performance, easy scaling to large parallel jobs, and integrated tooling that simplifies job management, pre‑processing, and post‑processing within Alibaba Cloud's EHPC ecosystem.