Oracle unveiled a high-performance computing (HPC)-heavy hardware roadmap today alongside a new partnership with service provider Rescale. The roadmap will see Oracle deploy Intel Ice Lake, AMD Milan, and Ampere Altra CPUs, as well as Nvidia A100 GPUs across Oracle Cloud Infrastructure (OCI).

According to Karan Batta, VP of product management at OCI, the hardware roadmap was informed by customer demand rather than the competition from the likes of Amazon Web Services or Microsoft Azure and reflects a recent push into the HPC market. Oracle recently announced two major customer wins with Nissan Motors and Altair Engineering, both of which will move on-premises workloads to Oracle's cloudn .

Meanwhile, a partnership with Rescale will pair the service provider's HPC-ready applications and support with OCI's cloud resources. Rescale currently hosts 450 applications running in OCI.

The cloud provider plans to deploy Intel's third-generation Scalable chips based on the forthcoming Ice Lake microarchitecture. Starting in early 2021, these chips will begin replacing Oracle's X7 HPC instances and are expected to deliver 30% higher performance in workloads like crash, computational fluid dynamics, and electronic design automation simulations.

Oracle also plans to offer Intel-based systems running on bare metal, which will enable customers to build large clusters using the cloud provider's remote direct memory access (RDMA)-enabled cluster network.

Further enhancing the cloud provider's HPC status, Oracle is among the first to deploy Nvidia's Ampere-based A100 GPUs with instances available as early as Sept. 30, in the U.S., Europe, Middle East, Africa, and Asian Pacific regions. In bare-metal configurations, Oracle claims customers will be able to configure up to 512 A100 GPUs in a single cluster, enabling large scale AI training or HPC workloads.

According to Batta, Oracle is approaching GPU acceleration a little differently than its competitors. "No. 1, you'll be able to get bare metal, which you can't get anywhere else," he said. "Second, you'll be able to get double the memory on an on-prem system."

Oracle's implementation pairs up to eight A100 with 2 TBs of memory,  compared to the 1 TB available in Nvidia's prepackaged DGX rack unit.

"The reason we do that is because a lot of our large retail customers on OCI actually prefer to run large graph workloads or large accelerated database workloads," Batta said.

AMD's upcoming Milan-based EYPC chips will power Oracle's E4 elastic compute instances beginning early next year.

Oracle's elastic compute instances enable customers to customize the core count and memory to meet their specific use cases. For example, a customer could create an instance with two virtual CPU cores and 12 GB of memory.

"The cool thing here is you'll be able to pick hardware architecture," Batta said. "It truly becomes an á la carte platform."

Oracle claims these instances will provide customers with superior price-to-performance ratios compared to competing cloud providers.

The E4 instances will complement Oracle's existing E3 instances, which are built on AMD's current generation Rome architecture.

Early this year Oracle announced plans to leverage Arm-based chipmaker Ampere's 80-core Altra series chips. Oracle claims these chips offer the best price-performance compared to any x86-based compute instance, at least on a per-core-basis. These chips will be available in bare-metal configurations with up to 160 cores running at 3.3 Ghz or as part of the company's flexible compute family of instances.

According to Batta, Oracle sees the adoption of Arm in the cloud as an opportunity to capitalize on the growing number of developers working on projects in mobile and edge that already have intimate knowledge of the architecture.

Oracle also aims to lower the barrier to entry for Arm-based processors by making them available on the company's always-free tier.