Improved Arm Server Price-Performance For HPC

The availability of Amazon EC2 Hpc7g instances with the AWS Graviton3E and Elastic Fabric Adapter (EFA) is opening new opportunities in key areas:

• Manufacturing
• Aerospace
• Automotive engineering
• Weather prediction

The new AWS EC2 instance types have AWS Graviton3E’s 64 Arm Neoverse V1 cores and 8 channels of DDR5 memory. This is alongside the AWS Nitro v5 card with EFA delivering 200GBps networking. The platform has serious HPC credentials!

Key ISV software packages in these HPC domains now support Arm architecture. So, Blue Chip corporations that rely on either commercial or open-source software are rapidly able to evaluate and transition their workloads to Arm. Compelling and efficient scale-out performance to thousands of cores across is demonstrated in the AWS Hpc7g deep-dive blog.

They looked at commercial and open-source CFD software in aerospace and automotive:

• OpenFOAM
• Ansys Fluent
• Ansys LS-DYNA
• Siemens StarCCM+
• GROMACS from molecular dynamics
• WRF from numerical weather prediction.

AWS’ own results show Hpc7g offering up to 70% better performance and almost 3x better price-performance compared to previous generation AWS Graviton instances for compute-intensive workloads.

Here we take similar workloads here and show the cost advantage of migrating them from x86 to Arm. We use the Hpc7g Arm architecture instances versus the existing HPC-specific Hpc6a instance type (AMD Milan).

We test these popular HPC applications:

• Aerodynamics: OpenFOAM with a 5.4 million element motorbike mesh
• Molecular Dynamics: GROMACS with the Bench RIB 12M atom peptides in water test case
• Weather Prediction: the CONUS 12.5KM model with WRF 4.2.2
• Crash Simulation: SkyCab drop test of 5.1M elements with OpenRADIOSS

We report each test with the best results found from the compiler and MPI combinations supported by the software package (Arm Compiler for Linux 23.04, GNU Compiler 12.2.0, AMD Optimizing C/C++ Compiler 4.0).

Execution time in CPU hours

In terms of elapsed CPU core hours, the AMD Milan instances are clocked at a higher frequency. The AWS Graviton3E cores, at 2.6GHz, are balanced for the DDR5 4800MHz memory and this suits the typical HPC workload.

The results are comparable in performance for molecular dynamics and weather prediction. The higher CPU frequency in AMD Milan shows some benefit. However, we see considerable reduction in total core hours for aerodynamics, which is a memory bound application. There are benefits from the 8xDDR5 memory controllers across the 64 Arm Neoverse V1 cores found in Graviton3E.

Cost saving

Now we turn to the execution cost and the savings from using Hpc7g over Hpc6a. Across every application, price-performance is better with Hpc7g. In automotive and aerospace (crash sim and aerodynamics), Hpc7g reduces cost by 18-25% over Hpc6a. The savings do not come at the expense of time to solution. This shows in AWS Hpc7g deep-dive scalability and our earlier execution time chart above. Graviton3 already demonstrates up to 60% greener compute through a reduction in energy use compared to other instances of similar compute performance. The arrival of Hpc7g is a competitive edge for industries that rely on HPC as Arm-based AWS Graviton3E processors enable them to achieve more within budget and improve end-to-end simulation and validation times.