AWS recently announced new bare metal instances for Amazon EC2. The new M5n, M5dn, R5n, and R5dn instances are already available and can utilize up to 100 Gbps of network bandwidth.
The new bare metal instances provide applications with direct access to the Cascade Lake processors and memory resources of the underlying server. These instances target workloads that require access to the hardware features, such as Intel VT-x and applications that need to run in non-virtualized environments for licensing requirements. Bare metal instances also make it possible to run virtualization secured containers such as Clear Linux Containers and, compared to virtual machines in a multi-tenant environment, avoid the "noisy neighbor" problem.
All the new bare metal instances, part of the general purpose and memory optimized EC2 families, offer higher network bandwidth and packet processing compared to the M5 and R5 instances, with 100 Gbps of network bandwidth and Elastic Fabric Adapter (EFA) for HPC/ML workloads. EFA is an optional EC2 networking feature to run applications requiring high levels of inter-node communications at scale on AWS. Based on a custom-built operating system, EFA bypasses hardware interfaces to enhance inter-instance communications. The improved network performance provides as well better data transfer to and from managed services like Amazon S3, reducing the data ingestion time for applications. M5dn and R5dn instances have local NVMe-based SSDs and provide block-level storage coupled to the instance.
The M5n, M5dn, R5n, R5dn instances were originally introduced in 2019 as EC2 virtual machines but were not available as bare metal hardware. The AWS announcement adds more technical details on the hardware and use cases:
M5n, M5dn, R5n, R5dn instances are powered by custom second-generation Intel Xeon Scalable processors (Cascade Lake) with sustained all-core turbo frequency of 3.1 GHz. They also provide support for the new Intel Vector Neural Network Instructions (AVX-512 VNNI) which will help speed up typical machine learning operations like convolution, and automatically improve inference performance over a wide range of deep learning workloads.
AWS introduced their first bare metal instances at re:Invent 2017 and added more options in the following years: in 2019 the cloud provider announced bare metal Arm instances, powered by Arm-based AWS Graviton processors, and at the end of last year added EC2 Mac instances.
AWS is not the only cloud provider offering bare metal servers. Google Cloud has the Bare Metal Solution, dedicated hardware to run specialized workloads with low latency, and Azure has Azure Dedicated Host, physical servers to host Azure VMs for Windows and Linux. IBM Cloud Bare Metal Servers is the single tenancy offer from IBM.
The new bare metal instances are available as on-demand, reserved or spot capacity. On-Demand pricing starts at $6.283 per hour in the region US East for a m5n instance, the cheapest of the four new offers.