High-performance computing architecture combines compute, memory, storage, and networking resources for demanding workloads.
High-performance computing (HPC) processes data and performs complex calculations quickly. HPC supports demanding artificial intelligence (AI), machine learning (ML), simulation, and electronic design automation (EDA) workloads. Cloud deployment can extend on-premises HPC capacity with compute and storage resources that can be allocated as workloads change.
HPC architecture combines compute resources with the memory, storage, interconnects, and software needed to run complex workloads. In the cloud, teams can provision resources for bursts in demand rather than sizing on-premises infrastructure for peak use at all times. Fine-tuning the architecture to the workload helps balance turnaround time, performance, and cost.
What Are the Design Principles for Cloud HPC Architecture?
Dynamic architecture
Design capacity to grow or shrink with natural workload cycles. Bursty projects can benefit from elastic, pay-as-you-go resources rather than systems that are idle between demand peaks.
Data visibility
Understand the origin, size, velocity, and updates of the data. Data services, analytics, and visualization can inform decisions about overall performance and cost.
Collaboration
Plan which tools, scripts, methods, results, and data will be shared, and with whom. This is especially relevant for projects spanning teams or geographic locations.
Cloud-native design
Consider cloud-native deployment patterns instead of simply copying an on-premises environment. Automated cluster deployment can help teams retire one compute cluster and launch another as requirements change.
Workload testing
Test realistic workloads before migration. HPC applications vary in memory, CPU, network behavior, algorithms, and model complexity, so generic benchmarks may not reliably predict their performance.
Cost versus time
Optimize less time-sensitive workloads for cost, potentially using spot or preemptible instances. Prioritize execution time for urgent workloads and account for provisioning, data staging, and queue time when comparing options.
How Does HPC Architecture Support EDA Workloads?
EDA workloads often require a compute cluster, a job scheduler, and a high-performance shared file system. Cloud resources can support parallel EDA jobs and help teams respond to changing compute demands. The architecture should be selected and tested against the characteristics of each workload, including the need for shared data, memory, and interconnect performance.
When matched to the workload, cloud-based HPC architecture can support compute-intensive AI, simulation, and chip design tasks without requiring every peak of demand to be met by owned infrastructure. [1]
Synopsys Cloud combines HPC infrastructure with on-demand access to electronic design automation (EDA) software licenses. Its cloud-based EDA platform and pre-optimized infrastructure configurations help teams deploy design and verification workloads while retaining established workflows.
Synopsys Cloud supports compute-intensive EDA projects with flexible access to software and cloud resources. Explore Synopsys Cloud to learn more about cloud-based chip design and verification.