AWS launches Graviton5-based EC2 C9g & C9gd instances

Amazon Web Services has made its Amazon EC2 C9g and C9gd instances generally available. The new compute-optimised instance families use AWS Graviton5 processors.

The launch gives AWS customers two new options for compute-heavy workloads. C9g is aimed at applications that rely on Elastic Block Store, while C9gd adds local NVMe SSD storage for tasks that need low-latency local access.

According to AWS, C9g delivers up to 25% higher performance per vCPU than the previous C8g generation. AWS also said the new line offers faster DDR5 memory, a larger L3 cache, and higher packet-processing performance than Graviton4-based instances.

C9gd adds local storage and, AWS said, offers up to 30% higher storage performance than the previous generation of local storage instances. Users of Graviton5-based instances with NVMe instance store volumes can also access detailed I/O statistics, including latency histograms by I/O size with up to one-second granularity, through Amazon CloudWatch or nvme-cli at no extra charge.

Range and sizes

The two instance families are available in 11 sizes, from medium to 48xlarge, as well as a bare metal option. At the top end, the 48xlarge and metal-48xl configurations offer 192 vCPUs and 384 GiB of memory.

AWS said the largest size provides up to 100 Gbps of network bandwidth and up to 72 Gbps of EBS bandwidth. Across the range, the families deliver, on average, up to 15% higher network bandwidth and 20% higher EBS bandwidth than the previous generation.

For C9gd, local storage ranges from 59GB in the medium configuration to three 3,800GB NVMe SSDs in the metal-48xl option. Smaller sizes offer up to 15 Gbps of network bandwidth, while larger configurations increase to 17 Gbps, 25 Gbps, 34 Gbps, 50 Gbps, and 100 Gbps.

Workload focus

AWS is targeting the new instances at high-performance computing, batch processing, gaming, video encoding, scientific modelling, distributed analytics, CPU-based machine learning inference, and ad serving. It also highlighted real-time analytics and what it described as agentic AI workloads, where CPU processing and concurrent environments are important.

Customers that need local scratch space for simulations, temporary caches for machine learning inference, or local buffers for ad-serving engines may prefer C9gd, AWS said. C9g is positioned for workloads that can rely on EBS for storage, including batch jobs, video encoding pipelines, and distributed analytics.

Both families support Instance Bandwidth Configuration, which lets users shift bandwidth allocation between EBS and Amazon VPC networking by up to 25%. They also support ENA Express, attachment of up to 128 EBS volumes on virtual instances, and purchasing models including Savings Plans, On-Demand, Spot Instances, Dedicated Instances, and Dedicated Hosts.

Isolation engine

The launch also marks the first compute-optimised EC2 instances to include the AWS Nitro Isolation Engine. AWS described it as a new Rust-written component of the Nitro Hypervisor that enforces isolation between virtual machines.

According to AWS, the Nitro Isolation Engine mediates access to virtual machine memory, CPU register state, and I/O devices through a limited set of APIs. The technology is part of the wider AWS Nitro System.

Initial availability covers US East in Ohio and North Virginia, US West in Oregon, and Europe in Frankfurt. The instances can be launched through the management console, command-line tools, and software development kits, AWS said.

No pricing details were included in the announcement.