Oracle Corp. today announced plans to upgrade its Oracle Cloud Infrastructure E5 instances to use the latest generation of Advanced Micro Devices Inc. EPYC processors.

The company will continue to offer the option of flexible instances that allow customers to allocate cores and memory as needed. The new EPYC processors provide about a 33% performance boost per core, 50% faster memory bandwidth and 50% more cores on bare metal instances compared with the previous generation at no additional cost, said Leo Leung, vice president of products and strategy at Oracle. Three customizable options will be available.

E5 Standard instances are aimed at the least-demanding uses such as web and application servers, enterprise applications and application development. Offered in both bare-metal and preemptible virtual machine configurations, the instances use external block storage, which is slower but less expensive than the local storage used in high-performance computing configurations.

Preemptible VM’s are intended for short-term usage and can be reclaimed by the cloud provider at any time when capacity is needed elsewhere in exchange for significant customer discounts.

High-performance instances

E5 Dense-IO instances are intended for use with large databases, big data workloads and applications that require speedy local storage. They provide 50% more storage capacity and 63% better performance than prior E4 Dense-IO instances. Local storage delivers roughly 10 times lower latency than standard block storage, Leung said.

E5 HPC instances are intended for complex mathematical and scientific problems like artificial intelligence training and weather forecasting. Offered only in bare-metal configurations, HPC instances come with optional cluster networking that combines the processing power of multiple HPC instances into a single virtual supercomputer. The HPC instances deliver roughly 40% better price performance than did previous HPC generations.

Cluster networking is based on remote direct memory access, a technology that has been long used in supercomputers to enable memory to be shared across computers without involving the CPU or operating system. RDMA achieves latency of as little as two microseconds, Leung said. “We can combine thousands of machines together so they are essentially one machine,” he said.

OCI provides what it says is the unique ability for customers to customize their deployments down to the number of CPU cores and storage capacity they need. Oracle said this option is saving customers over $40 million a year compared with the more rigid configuration options that are typical of public cloud environments.

“If you look at every other cloud provider you can only get machines that are one, two, four, eight or 16 cores,” Leung said. “A large part of our user population is using three cores or nine cores. The savings is the difference between using nine cores, for example, and paying for sixteen cores.” He said that customers can achieve ancillary benefits by taking advantage of per-core software licensing.

E5 instances can be configured with clustered file systems for such use cases as large databases, AI model training, financial analysis, video rendering and simulations. The new options will be rolled out during July and August. Existing customers will need to go through a structured migration process that Leung described as “not complicated.”

He declined to compare the price/performance of AMD instances with those from rival Intel Corp. AMD has increasingly targeting the HPC market in recent years and emphasized the greater core density of its CPUs compared with Intel’s.

Photo: Flickr CC