The convergence of simulation, artificial intelligence (AI), and high-performance computing (HPC) is bringing exciting new opportunities to automotive engineering. From design to manufacturing to supply chain management, Altair’s solutions are helping the world’s largest manufacturers and suppliers accelerate development, enhance safety, optimize vehicle architectures, and reduce costs.

However, manufacturers are grappling with new challenges as they embrace new design and simulation techniques. These include heightened competition, stringent emission and fuel efficiency requirements, and the need for innovative simulation methods to address issues like electromagnetic compatibility, battery life, and thermal runaway. 

Infrastructure for engineering simulation poses a particular challenge. Manufacturers need the capacity to quickly deploy complex HPC environments and run high-fidelity simulations faster while minimizing simulation costs and their carbon footprint.  

This article shares recent benchmarks conducted by AMD, Google, and Altair – and their breakthrough results on Google Cloud powered by the latest 4th Gen AMD EPYC™ processors. 

About the Benchmarks

Automotive engineering involves a wide range of simulations in disciplines that include finite element analysis (FEA), computational fluid dynamics (CFD), crash simulation, and modeling of electromagnetic compatibility and interference (EMC/EMI). 

To assess the performance of the latest Google Cloud C3D compute instances, AMD ran four benchmarks using Altair solvers and standard benchmarks to simulate automotive workloads. The applications tested were as follows:

• Altair® AcuSolve®, the general-purpose Navier-Stokes implicit solver offered in Altair CFD™, is an advanced modeling solution that supports accurate simulation for a full range of flow, heat transfer, turbulence, and non-Newtonian simulation problems. The benchmark used Altair’s standard Impinging Nozzle model, which simulates jets commonly used for convective cooling applications.
• Altair® Feko® is a high-frequency electromagnetic simulator used for various applications, including optimizing wireless connectivity, ensuring electromagnetic capability and compliance, and performing radar cross-section and scattering analysis — capabilities critical for simulating sensors in autonomous vehicles, LiDAR modeling, and advanced driving assistance systems (ADAS). Two standard models were used in the benchmark.¹
• Altair® OptiStruct® is a proven structural solver that supports linear and nonlinear analysis across a variety of statics, dynamics, vibration, acoustics, and multiphysics disciplines. In this test, a standard Altair Engine Assembly model was used.
• Altair® Radioss® is widely used to simulate highly nonlinear problems under dynamic loadings to improve vehicle crashworthiness, safety, and manufacturability. Three simulated crash tests used standard Altair models, including a Chrysler Neon (neon), a Ford Taurus (t10m), and an electric Toyota Venza collision involving the battery pack (venza battery).

In each case, the simulations were run three times, and the average results were tabulated, comparing the performance of Google’s latest standard C3D instances with that of previous-generation C2D instances. The HPC cluster was deployed using the Cloud HPC Toolkit offered by Google Cloud, which makes it easy for users to deploy HPC environments. The simulations were run with Simultaneous Multithreading (SMT) disabled for the instances.  

Compelling Results in the Google Cloud

Simply shifting workloads to the latest C3D instances resulted in an average performance improvement of ~2.23x! Improvements ranged from ~1.74x running the static engine block simulation with OptiStruct to ~2.53x with Feko.  The results for each Altair solver are illustrated below:

To put this in perspective, a simulation that previously took one hour to run could be completed in under 27 minutes, literally doubling engineering productivity.² With this additional performance, organizations have the flexibility to explore additional design scenarios or run higher-fidelity simulations within the same time window. 

C3D instances, powered by the latest 4th Gen EPYC™ processors, deliver over 3x the virtual core count and memory compared to C2D instances supporting larger CAE models.³ For Altair solvers supporting MPI, C3D instances provide up to 200 Gbps of tier 1 internal egress bandwidth and reduced latency between internal instance IP addresses.⁴ The table below provides specifications of the cloud instances tested.

The AMD EPYC™ Advantage

The latest AMD EPYC processors combine high core counts, large memory capacity, extreme memory bandwidth, large cache sizes, and massive I/O to deliver exceptional performance for automotive applications. For CAE customers, this means faster, more thorough simulations of larger, more complex models, helping deliver a significant competitive advantage. 4th Gen AMD EPYC processors offer multiple improvements over the previous generation:

• Up to 96 cores per processor, delivering 1.45x the floating-point throughput of alternative CPUs.⁵
• A faster Infinity Fabric, delivering 2x the speed of the previous generation.⁶
• Greater I/O capacity than competing processors.⁷
• Advanced chip-level security enhancements (SME, SEV-ES, SEV-SNP).⁸

The latest EPYC processors are also energy efficient, delivering approximately 1.8x the throughput per watt based on the SPECpower_ssj2008 benchmark compared to competing processors.⁹ This means that manufacturers can achieve dramatic performance and energy efficiency gains by selecting AMD EPYC for on-premises deployments or using Google’s EPYC-powered C3D cloud instances.

Cloud Computing Brings Clear Benefits

While many organizations choose to run on-premises CAE environments, running workloads in the Google Cloud can offer several advantages:

• Instant access to the latest state-of-the-art CPUs and GPUs for simulation and model training.
• Opportunities to reduce costs by paying for resources only when required.
• No need to manage complex on-premises HPC infrastructure.
• Enhanced power efficiency to meet sustainability goals.

In addition to taking advantage of Google’s best practices for securing applications in the cloud, manufacturers can protect sensitive intellectual property (IP) with Google Cloud Confidential Computing. AMD EPYC CPU- powered Google Cloud instances take advantage of AMD Secure Encrypted Virtualization (SEV) to help safeguard data in VMs during processing. Data is encrypted in silicon such that even the hypervisor cannot access sensitive data handled by hosted VMs.  

Increasingly, automotive manufacturers have sustainability goals. While designing more fuel-efficient and environmentally friendly products is imperative, most also have stated goals to reduce carbon dioxide (CO₂) emissions. Running simulations in the cloud shifts emissions from scope 2 to scope 3 while enabling manufacturers to take advantage of the cloud's massive scale and efficiencies.¹⁰ Google aims to achieve net-zero emissions by 2030, helping organizations achieve their sustainability targets.¹¹

Highlighting Google’s commitment to sustainability, Google is sponsoring a new Sustainable Computing category in this year’s Altair Enlighten Award, which highlights innovations in sustainable, lightweight design for the automotive industry. Winners will be announced in August 2024.

Getting Started with Altair Solvers in the Google Cloud

Google provides multiple solutions for provisioning HPC infrastructure in the cloud. The benchmark environments described above were deployed using Cloud HPC Toolkit, an open-source solution offered by Google Cloud that makes it easy to deploy scalable HPC environments based on standard templates.

Another way to get started is to sign up for a free Altair One® account. Altair One provides seamless access to your software, scalable Google Cloud resources, and data from any device. Users can visit the Altair Marketplace and access nearly 200 applications from Altair and third parties. 

<sup>1. The Altair® Feko® benchmarks involved analyzing a ship's radar cross-section (RCS) using multilevel fast multipole (MLFMM) analysis and a second test with a sparse-approximate inverse (SPAI) preconditioner.

2. A simulation that normally takes one hour (3,600 seconds) would require 3600/2.23/60 = 26.9 minutes to complete with a 2.23x performance improvement.

3. The c3d-standard-360 instance uses 2 x AMD EPYC™ processors, each with 90 cores, providing 360 threads or vCPUs compared to 112 threads or vCPUs for the test c2d-standard-112 instances; a ~3.2x increase.

4. Google Cloud Tier 1 networking is an option available for selected VMs that provides enhanced network transfer speeds and reduced latency for HPC, ML, and DL applications. To achieve optimal bandwidth, instances must be in the same cloud zone and attached to the same VPC.

5. SP5-009D: SPECrate 2017_fp_base based on published scores from spec.org as of Jan 11, 2023. Configurations: 2P AMD EPYC 9654 (1480 SPECrate 2017_fp_base, 192 total cores, spec.org/cpu2017/results/res2022q4/cpu2017-20221024-32605.html) is 1.45x the performance of published 2P Intel Xeon Platinum 8490H (1020 SPECrate 2017_fp_base, 120 total cores, spec.org/cpu2017/results/res2023q1/cpu2017-20221206-33040.html). SPEC®, SPEC CPU®, and SPECrate® are registered trademarks of the Standard Performance Evaluation Corporation. See spec.org for more information.

6. See description of the AMD Infinity Architecture in the 4th Gen AMD EPYC Processor Architecture whitepaper on page 7.

7. EPYC-035A: One AMD EPYC 9004 CPU supports 128 PCIe 5 lanes plus up to 8 PCIe 3 lanes. One 4th Gen Intel Xeon Scalable CPU supports up to 80 lanes of PCIe 5 per ark.intel.com/. EPYC 9004 series offers 128 ÷ 80 = 1.6x more PCIe 5 lanes.

8. For details, see https://www.amd.com/en/products/processors/server/epyc/infinity-guard.html

9. SP5-011C: SPECpower_ssj2008 comparison based on published 2U, 2P Windows results as of 11/10/2022. Configurations: 2P AMD EPYC 9654 (30,602 overall ssj_ops/W, 2U, spec.org/power_ssj2008/results/res2022q4/power_ssj2008-20221204-01204.html) vs. 2P Intel Xeon Platinum 8480+ (16,653 overall ssj_ops/W, 2U, spec.org/power_ssj2008/results/res2023q1/power_ssj2008-20221207-01216.html). 2P AMD EPYC 7763 (23,505 overall ssj_ops/W, 2U, spec.org/power_ssj2008/results/res2021q2/power_ssj2008-20210324-01091.html) shown at 1.4x for reference. SPEC® and SPECpower_ssj® are registered trademarks of the Standard Performance Evaluation Corporation. See spec.org for more information. 30,602 ssj_ops/W / 16,653 ssj_ops/W = 1.84 energy efficiency improvement.

10. For an explanation of Scope 1,2 and 3 emissions, see the EPA Center for Corporate Climate Leadership guidance at https://www.epa.gov/climateleadership/scope-1-and-scope-2-inventory-guidance.

11. https://sustainability.google/operating-sustainably/net-zero-carbon/</sup>