AWS Amazon EC2 M9g and M9gd Instances Powered by Graviton5 Processors Deliver Massive Performance Gains for AI and Enterprise Cloud Workloads
Amazon Web Services has publicly announced the general availability of their new generation of virtual machines, the Amazon EC2 M9g and M9gd instances. This new infrastructure, which is powered by their proprietary designed Graviton5 processor, is said to provide more computational power for demanding AI centric applications such as real time reasoning, automatically generated code, and multi step task orchestration. This release, which follows up a previous preview, shows that AWS is providing major hardware upgrades for businesses which are attempting to scale up their local and distributed AI workload operations.
The Graviton5 processor comes with some huge physical improvements over previous generation chips, featuring 192 processing cores on a single silicon die thanks to advanced 3 nm manufacturing. Furthermore, this chip can operate with DDR5 8800 memory, the fastest memory that exists currently for cloud computing. Also included is PCIe Gen 6 support as well as an L3 cache 5x the size of its predecessors which allows each of the processor cores to use 2.6x more high speed cache than before. This improves inter core communication latency by up to 33% and increases data bandwidth for dense workload scaling.
Performance benchmarks for different EC2 offerings and configurations have shown that the new machines perform remarkably better than its predecessors. The company has announced two distinct offerings to fit a variety of corporate infrastructure needs. Standard EC2 M9g instances provide up to 25% more computing power over the previous generations of chips. The specific improvements in the architecture include a 35% increased speed in web application usage, 35% more performance on machine learning inference, and 30% increased database performance. EC2 M9g instances also perform up to 15% better than previous machines on the network and provide 20% better Amazon Elastic Block Store performance on average across their configurations.
The high speed local storage option from the new EC2 machines, the M9gd, come with up to 11.4TB of local NVMe SSD storage, which offers a 30% improvement in input/output operations per second than previous generations of silicon. Both instance types are built on the 6th generation of AWS Nitro System infrastructure which provides the Nitro Isolation Engine which uses formal verification to provide complete isolation of virtual machines. This provides a huge leap in security and data protection on the cloud.
Enterprises around the world have begun utilizing these chips to meet their demands. The cloud provider has been collaborating with Meta, who have pledged to commit millions of its Graviton5 cores to meet its growing AI workload demands and other companies such as Uber and Snowflake are making plans to move their workload to the new architecture. Adobe has also adopted the new chips for its real time video streaming services, and multiplayer gaming giant Epic Games has found the lower latency beneficial in its gaming services. On the cloud provider itself, Pinterest is operating its platform that services over 500,000,000 active users per month on their own infrastructure that utilizes the Graviton chips and Formula 1 utilizes their chips for data telemetry when processing vehicles on the race track at 350 kilometers per hour.
In testing it has been found that these chips can even produce a performance gain of up to 25% on the real world search workloads being used at Airbnb over other competitive architecture. The engineer at Airbnb that identified this gain also stated that performance of their current infrastructure on the Graviton5 chips jumped 20% compared to the previous generation of chips. This gain was due to a large performance decrease in latency on their most critical user services. Similarly Atlassian, the company behind Jira and Confluence has recently migrated their infrastructure of over 3000 EC2 instances to use the new Graviton chips. One of their site reliability engineers identified a 30% performance increase in Jira on the M9g machines when testing it on the M9g instance type with 20% less latency than their previous machines.
Both the semiconductor market and enterprise software vendors are seeing impressive gains on the chips. Siemens Digital Industries Software has now confirmed that its Calibre design platform has been optimized to work with Arm based Graviton processors, with its senior vice president stating they saw a 30% increase in performance when working with the Graviton5 chip and had already seen a 30% cost decrease using Graviton4 machines. Synopsys also came out with a similar claim, stating their design automation tools worked 35% faster on the Graviton5 chips.
Database management systems saw some of the largest performance gains on the chips with SAP providing benchmark reports of 35% to 60% improved performance in transactional query performance when running their SAP HANA Cloud on the Graviton5 machines.
