Governing and Running HPC/AI Workloads in the Cloud

Hyperion Research has followed the growth of accelerators in HPC for more than 10 years. Accelerators had an initially slow HPC adoption curve but have recently experienced amplified growth. Within the HPC accelerator market, GPUs, particularly those from Nvidia, have dominated. Although originally designed as gaming processors for graphics rendering, GPUs were found to be well suited to a range of HPC applications, plus critical underlying AI functions, such as matrix multiplication, driving impressive performance gains. As a result, many HPC sites across a broad range of verticals are now using accelerators to speed up a larger portion of their increasingly AI-based workloads.

HPC cloud adoption has grown aggressively over the last few years as cloud providers have recognized the HPC market as a high value target and have brought HPC-specific capabilities to their platforms to entice users. This has resulted in growing number of HPC users that are now using external clouds for parts of their HPC workload portfolio.   Hyperion Research has followed the growth vector of cloud adoption for many years and conducted a number of broad studies to better understand the underlying trends of cloud usage. This research has shown that users are contemplating different approaches of cloud adoption including what types of applications to send to the cloud; how much to send to the cloud; and ultimately whether to run a specific workload in the cloud or not.   As the performance capabilities of the cloud for HPC workloads improve, many of the barriers to adoption have been reduced, but some remained consistent barriers despite the technological advancements. These adoption inhibitors may be offset in many situations by the corresponding drivers and benefits. This will lead to increased cloud usage for many HPC applications and workloads in both traditional modeling/simulation as well as artificial intelligence.