High Performance Computing

2 posts

aws3 min readCurated summary

Amazon EC2 C9g and C9gd instances powered by AWS Graviton5 processors are now available | Amazon Web Services

Amazon EC2 C9g and C9gd instances, powered by AWS Graviton5, are now generally available for compute-intensive workloads. They provide up to 25% better performance per vCPU than C8g, faster DDR5 memory, larger caches, and improved networking and EBS bandwidth. C9gd adds local NVMe SSD storage, making it suitable for workloads requiring both high CPU performance and low-latency temporary storage. ## Graviton5 Performance Improvements - Up to 25% higher performance per vCPU than previous-generation C8g instances. - DDR5 memory running at 8800 MT/s, described as the fastest memory available in a cloud processor instance. - Five times more L3 cache than Graviton4-based instances. - Up to three times higher packet-processing performance than Graviton4. - Benefits include faster in-memory analytics, higher throughput, and more responsive real-time applications. ## C9g and C9gd Workloads - C9g is designed for compute-heavy applications using Amazon EBS, including: - Batch processing - Video encoding - Distributed analytics - CPU-based machine learning inference - Agentic AI workloads - C9gd adds local NVMe SSD storage for: - HPC simulation scratch space - Machine learning inference caches - Ad-serving buffers - Other low-latency temporary-storage use cases - C9gd delivers up to 30% higher local storage performance than the previous generation. ## Networking, Storage, and Configuration - Available in 11 sizes from medium through 48xlarge, plus bare metal. - Up to 15% higher network bandwidth and 20% higher EBS bandwidth on average compared with the prior generation. - The largest instances provide up to 100 Gbps networking and 72 Gbps EBS bandwidth. - Instance Bandwidth Configuration allows up to 25% of bandwidth to be shifted between EBS and VPC networking. - Support includes ENA Express, up to 128 EBS volumes, and On-Demand, Spot, Savings Plans, Dedicated Instances, and Dedicated Hosts. - NVMe-equipped instances expose detailed I/O statistics, including latency histograms by I/O size at one-second granularity through CloudWatch or `nvme-cli`. ## Nitro Isolation Engine - C9g and C9gd are the first compute-optimized EC2 instances to use the AWS Nitro Isolation Engine. - The Rust-based Nitro System component isolates virtual machines by mediating access to memory, CPU register state, and I/O devices through a minimal API set. - AWS provides additional technical documentation covering the engine and its formal verification results. ## Availability - The instances are available in US East (Ohio and Northern Virginia), US West (Oregon), and Europe (Frankfurt). - They can be launched through the AWS Management Console, CLI, or SDKs, with more regions planned. For compute-intensive workloads, C9g is the general-purpose choice, while C9gd is preferable when fast local NVMe storage is also required.

Read original(opens in new tab)
aws2 min readCurated summary

Amazon EC2 Hpc8a Instances powered by 5th Gen AMD EPYC processors are now available | Amazon Web Services

Amazon EC2 Hpc8a instances are now generally available for tightly coupled, compute-intensive HPC workloads. Powered by 5th Gen AMD EPYC processors reaching 4.5 GHz, they provide up to 40% more performance, 42% higher memory bandwidth, and 25% better price-performance than Hpc7a instances. AWS targets applications such as fluid dynamics, weather modeling, design simulations, and crash analysis. ## Instance Specifications - Available in a single `96xlarge` configuration: - 192 CPU cores - 768 GiB memory - 300 Gbps Elastic Fabric Adapter (EFA) networking - Uses a 1:4 core-to-memory ratio. - Customers can customize the number of cores at launch to better match workload requirements. - Simultaneous Multithreading (SMT) is disabled to maximize HPC performance. - Sixth-generation AWS Nitro cards handle virtualization, storage, and networking tasks separately from the CPUs. ## Supported HPC Services - Integrates with AWS ParallelCluster and AWS Parallel Computing Service (AWS PCS) for cluster creation and job submission. - Supports Amazon FSx for Lustre, offering sub-millisecond latency and throughput of up to hundreds of gigabytes per second. - High-bandwidth, low-latency networking is designed for workloads requiring extensive communication between compute nodes. ## Availability and Purchasing - Initially available in: - US East (Ohio) - Europe (Stockholm) - Offered through On-Demand Instances and Savings Plans. - Regional availability and future expansion can be checked through AWS Capabilities by Region. Hpc8a instances are best suited for organizations needing faster simulation results and efficient scaling across tightly coupled HPC workloads. Teams can launch them through the Amazon EC2 console and combine them with AWS cluster and storage services for a complete HPC environment.

Read original(opens in new tab)