The selection of processor (CPU), more specifically whether to prioritise cores over clock speeds, can be a crucial decision when designing a cluster
As a technical presales worker in the world of high-performance computing (HPC), the selection of processor (CPU), more specifically whether to prioritise cores over clock speeds, can be a crucial decision when designing a cluster. These choices can influence a cluster's performance differently, depending on the type of workload being utilised by a customer.
First, it’s important to understand what the terms “Cores” and “Clock Speed” mean:
From a high-level, HPC tasks are typically a mix between parallel and serial workloads. Parallelised workloads tend to benefit from more physical cores, whereas serial workloads tend to favour faster clock speeds. Understanding the definition and difference between parallelised and serial workloads is also important when selecting the right CPU for an HPC cluster, especially in fields such as engineering and higher education.
A parallelised workload is a task that can be divided into multiple, independent operations, thus allowing each task to be processed simultaneously. More cores will enable these tasks to be executed together. As the overall number of cores within a system increases, the workload at hand can scale effectively, allowing more complex or larger tasks to be handled at once.
Many engineering, machine learning, and scientific simulations are common examples of parallelised workloads within HPC. In these cases, more cores will allow for more computations to be processed at any given time, and for tasks that can be broken down into smaller operations, additional cores will enable a larger portion of the workload to be handled at once. Engineering applications such as ANSYS, coupled with scientific applications like GROMACS and NAMD, are highly parallelised, designed to distribute computations across multiple cores or nodes rather than running serially.
On the other hand, a serial workload is a task that cannot easily be divided into independent sub-tasks and must be completed on a step-by-step basis. These types of workloads can benefit greatly from higher clock speeds, due to operations depending on the timings of the previous ones, meaning that the CPU needs to execute each step as quickly as possible before starting a new operation.
While HPC often focuses on parallel workloads, there are still instances where serial processing is necessary. Examples can include certain types of data analysis and data processing that require specific order, as well as legacy software or inherently sequential tasks that may not benefit from parallel processing.
In summary, parallel workloads will benefit greatly from higher core counts, allowing multiple operations to be executed simultaneously, while serial workloads will rely more on higher clock speeds to complete step-by-step tasks more efficiently.
As of writing, both AMD and Intel now offer processors with exceptionally high core counts compared to that of just a few years ago. The AMD EPYC 9965 features 192 cores, while the Intel 6980P offers 128 cores. As discussed, however, such high core count CPUs are not always the optimal choice for every workload. Lower core count processors from both vendors can provide vastly superior clock speeds.
In typical HPC environments, most clusters must support a mixture of workload types, ranging from highly parallel to largely serial, making it important to find the right balance between core count and clock speed when selecting the processors for an HPC cluster and optimising performance.
If you are interested in discussing your specific requirements,
please contact David or a member of the Presales Team at OCF on +44 (0)114 257 2200.