We recently invited Exabyte.io, a cloud-based, nanoscale modeling platform that accelerates research and development of new materials, to test the high-performance computing (HPC) hardware in Oracle Cloud Infrastructure. Their results were similar to the performance that our customers have been seeing and what other independent software vendors (ISVs) have been reporting: Oracle Cloud Infrastructure provides the best HPC performance for engineering and simulation workloads.
Exabyte.io enables their customers to design chemicals, catalysts, polymers, microprocessors, solar cells, and batteries with their Materials Discovery Cloud. Exabyte.io allows scientists in enterprise R&D units to reliably exploit nanoscale modeling tools, collaborate, and organize research in a single platform.
As Exabyte.io seeks to provide their customers with the highest-performing and lowest-cost modeling and simulation solutions, they have done extensive research and benchmarking with cloud-based HPC solutions. We were eager to have them test the Oracle Cloud Infrastructure HPC hardware. Exabyte.io ran several benchmarks, including general dense matrix algebra with LINPACK, density functional theory with Vienna Ab-initio Simulation Package (VASP), and molecular dynamics with GROMACS.
The results were impressive and prove the value, performance, and scale of HPC on Oracle Cloud Infrastructure. The advantage of Oracle Cloud Infrastructure's bare metal was obvious with LINPACK, throughput is almost double the closest cloud competitor and consistent with on-premises performance. Latency is even more interesting: the BM.HPC2.36 shape with RDMA provides the lowest latency at any packet size and is orders of magnitude faster than cloud competitors. In fact, for every performance metric that Exabyte.io tested on VASP and GROMACS, they saw Oracle's BM.HPC2.36 shape with RDMA (shown as OL in the following graph) outperform the other cloud competitors. Below is a great example of both performance and scaling of Oracle Cloud Infrastructure on VASP. When parallelizing over electronic bands for large-unit-cell materials and normalizing for core count, the single node performance of the BM.HPC2.36 exceeds it's competitors and then scales consistently as the cluster size increases. The BM.HPC2.36 runs large VASP jobs faster and can scale larger than any other cloud competitor.
Exabyte.io has provided the full test results on their website. Their blog concluded that "Running modeling and simulations on the cloud with similar performance as on-premises is no longer a dream. If you had doubts about this before, now might be the right time to give it another try."
By offering bare metal HPC performance in the cloud Oracle Cloud Infrastructure enables customers running the largest workloads on the most challenging engineering and science problems to get their results faster. The results that Exabyte.io has seen are exceptional, but the results are not unique among our customers. Spin up your own HPC cluster in 15 minutes on Oracle Cloud Infrastructure.