In today’s accelerated business environment, the foundation for successful HPC technology adoption begins with a well-defined HPC architecture. Depending on your organization’s workloads and computing goals, different HPC system designs and supporting resources are available to help you achieve productivity gains and scalable performance.
Designing HPC Systems
HPC architecture takes many forms based on your needs. Organizations can choose different ways to design HPC systems.
Parallel Computing Across Multiple Architectures
HPC parallel computing allows HPC clusters to execute large workloads and splits them into separate computational tasks that are carried out at the same time.
These systems can be designed to either scale up or scale out. Scale-up designs involve taking a job within a single system and breaking it up so that individual cores can perform the work, using as much of the server as possible. In contrast, scale-out designs involve taking that same job, splitting it into manageable parts, and distributing those parts to multiple servers or computers with all work performed in parallel.
As intensive workloads such as simulations, modeling, and advanced analytics become more commonplace, HPC systems are being designed to incorporate accelerators in addition to CPUs. These accelerators have introduced a wider, heterogeneous range of possible configurations that developers need to support.
Developers can use oneAPI cross-architecture programming to create a single codebase that can be used across CPU, GPU, and FPGA architectures for more productive and performant development. oneAPI can accelerate HPC innovation by removing the constraints of proprietary programming models, easing adoption of new hardware, and reducing code maintenance. oneAPI and Intel® oneAPI toolkits support existing open standards and languages that HPC developers need, including C++, SYCL, Fortran, OpenMP, MPI, and Python. Get more details on oneAPI and Intel® oneAPI toolkits.
Cluster Computing
High performance computing clusters link multiple computers, or nodes, through a local area network (LAN). These interconnected nodes act as a single computer—one with cutting-edge computational power. HPC clusters are uniquely designed to solve one problem or execute one complex computational task by spanning it across the nodes in a system. HPC clusters have a defined network topology and allow organizations to tackle advanced computations with uncompromised processing speeds.
Grid and Distributed Computing
HPC grid computing and HPC distributed computing are synonymous computing architectures. These involve multiple computers, connected through a network, that share a common goal, such as solving a complex problem or performing a large computational task. This approach is ideal for addressing jobs that can be split into separate chunks and distributed across the grid. Each node within the system can perform tasks independently without having to communicate with other nodes.
HPC Cloud Infrastructure
In the past, HPC systems were limited to the capacity and design that on-premises infrastructure could provide. Today, the cloud extends local capacity with additional resources.
The latest cloud management platforms make it possible to take a hybrid cloud approach, which blends on-premises infrastructure with public cloud services so that workloads can flow seamlessly across all available resources. This enables greater flexibility in deploying HPC systems and how quickly they can scale up, along with the opportunity to optimize total cost of ownership (TCO).
Typically, an on-premises HPC system offers a lower TCO than the equivalent HPC system reserved 24/7 in the cloud. However, an on-premises solution sized for peak capacity will be fully utilized only when that peak capacity is reached. Much of the time, the on-premises solution may be underutilized, leading to idle resources. However, a workload that can’t be computed because of a lack of available capacity can result in a lost opportunity.
In short, using the cloud to augment your on-premises HPC infrastructure for time-sensitive jobs can mitigate the risk of missing big opportunities.
To drive HPC innovation in the cloud, Intel works deeply with cloud service providers to maximize performance, apply technologies such as Intel® Software Guard Extensions (Intel® SGX) and Intel® Advanced Vector Extensions 512 (Intel® AVX-512), and simplify onboarding. Read more about our HPC cloud technologies and how they can help enhance results.
Selecting HPC Processors for Scalability and Performance
With our breadth of expertise in HPC technologies, Intel delivers the performance requirements for handling the most-demanding future workloads. Intel® Xeon® Scalable processors provide a highly versatile platform that can seamlessly scale to support the diverse performance requirements of critical HPC workloads. Our upcoming data center GPUs based on the Xe HPG microarchitecture will make an ideal complement to Intel® Xeon® Scalable processors, helping to drive performance even further.
Working with our ecosystem, Intel prioritized efforts in creating blueprints that help enable optimized HPC system designs. For validating performance requirements, Intel® Cluster Checker (part of the Intel® oneAPI HPC Toolkit) ensures that your HPC cluster system is intact and configured to run parallel applications with incredible portability for moving between on-premises and HPC cloud systems.
With Intel® CoFluent™ technology, you can speed up the deployment of complex systems and help determine optimal settings by modeling simulated hardware and software interactions.
A Breakthrough in HPC Memory
Memory is an integral component of HPC system design. Responsible for a system’s short-term data storage, memory can be a limiting factor to your workflow performance. Intel® Optane™ technology helps overcome these bottlenecks in the data center by bridging gaps in the storage and memory hierarchy, so you can keep compute fed.
High-bandwidth memory has demonstrated success when included in the GPUs used in HPC programming and will soon be available on x86 processors as well. In many cases, high-bandwidth memory technologies can accelerate codes that are memory bandwidth bound without code changes. It can also help reduce DDR memory costs.
Scaling Performance with HPC Fabric
To effectively scale smaller HPC systems, organizations need a high-performance fabric that is designed to support HPC clusters.
Intel® high performance networking (HPN) provides optimized performance using familiar and cost-effective Ethernet technologies. This gives cluster managers an end-to-end solution that covers their HPC and machine learning training needs. Existing popular HPC and AI middleware and frameworks, including oneAPI, can be used with Intel® HPN through the OpenFabrics Interfaces (OFI, also known as libfabric) and the Intel® Ethernet Fabric Suite.
An Easier Path to HPC Success
Intel provides the expertise to deeply understand HPC applications, architectures, and what an HPC system—whether on-premises, in the cloud, or hybrid—requires to help users produce results and maximize accomplishments. With HPC architecture based on a foundation of Intel® technologies, you can be ready to meet the HPC, exascale, and zettascale needs of the future.
Plus, our oneAPI toolkits are ready to help developers simplify their HPC programming efforts, allowing them to support more hardware types and maximize business results.