AWS Graviton5 and EC2 M9g Instances: Full Specs, Benchmarks, and What Changes

After months of preview, AWS Graviton5-powered Amazon EC2 M9g and M9gd instances became generally available on June 10, 2026, marking the broadest rollout yet for Amazon's custom Arm server silicon . The new instances pair a dense 192-core processor built on a 3nm-class TSMC process with the sixth-generation Nitro System, and they introduce the Nitro Isolation Engine—the first production cloud hypervisor backed by machine-checkable formal verification proofs .

The result is a general-purpose instance family that pushes past Graviton4 on nearly every axis while adding a security property no other major public cloud ships in production today: mathematical proof that customer workloads are isolated from one another and from AWS operators .

Graviton5 Processor Specifications

Graviton5 represents Amazon's most significant architectural leap since the Graviton family debuted. The headline numbers:

• Cores: Up to 192 Armv9 Neoverse V3 cores on a single die, eliminating the dual-socket NUMA topology that Graviton4 used to reach 192 cores (96 per socket)
• Process node: 3nm-class, widely attributed to TSMC
• Clock speed: 3.1 GHz on the Neoverse V3 cores
• L3 cache: 180 MB shared—a 5× increase over Graviton4's 36 MB, with 2.6× more L3 cache available per core
• Memory: 12 DDR5 controllers supporting up to DDR5-8800, the fastest DDR5 speed offered by any major cloud provider at launch
• I/O: PCIe Gen 6 support across eight controllers
• Inter-core latency: 33% lower than Graviton4, a direct consequence of the single-socket design

Moving from a dual-socket Graviton4 design to a monolithic Graviton5 die eliminates cross-socket communication overhead entirely. For workloads that spread across many cores—real-time inference pipelines, in-memory databases, or large-scale microservice fleets—the latency reduction alone can yield measurable throughput gains before any IPC improvements are considered.

Performance Gains Over Graviton4 (M8g)

AWS's published generational improvements are consistent across official sources, third-party analysis, and early customer benchmarks:

Compute and throughput:

• Up to 25% better general compute performance vs. M8g
• Up to 35% faster web application serving
• Up to 35% faster machine learning inference on CPU
• Up to 30% faster database workloads

I/O and bandwidth:

• Network bandwidth: higher across all instance sizes; largest sizes can double network throughput compared to equivalent M8g configurations
• EBS bandwidth: approximately 20% higher on average, with the largest instances again seeing the biggest proportional gains

Real-world customer results:

• ClickHouse observed a 36% performance boost versus M8g with zero code changes
• Honeycomb measured 36% better throughput per core compared to Graviton4 during a six-month A/B test of production observability workloads
• HubSpot deployed M9g for MySQL databases and saw query duration drop by up to 60%

These numbers align with the architectural changes. The 5× larger L3 cache reduces costly DRAM accesses, particularly for database and analytics workloads that traverse large working sets. The faster DDR5-8800 memory and PCIe Gen 6 I/O remove bandwidth bottlenecks that capped throughput on previous generations. And the shift to a single-socket design reduces the latency tax that scaled-out applications pay on NUMA architectures.

M9gd Variant: Local NVMe SSD Storage

For workloads that need high-speed ephemeral storage directly attached to the instance, AWS offers the M9gd variant. These instances layer local NVMe-based SSD block storage on top of the same Graviton5 compute platform, providing up to 11.4 TB of local NVMe SSD capacity with 30% higher IOPS than the previous generation's local storage offering .

The M9gd variant targets workloads like large-scale caching fleets, log processing pipelines, and real-time analytics engines where keeping data as close to the CPU as possible directly impacts query latency and throughput. The combination of faster cores, lower inter-core latency, and higher local storage IOPS makes the M9gd a natural fit for any workload that benefits from collapsing the storage-compute gap.

Positioning for Agentic AI Workloads

One of the more notable positioning shifts with Graviton5 is AWS's explicit targeting of agentic AI workloads—systems that perform real-time reasoning, code generation, and multi-step task orchestration using large language models and other generative AI techniques .

While GPU and accelerator instances dominate the training and large-batch inference conversation, agentic AI at scale creates a different compute pattern: continuous high-throughput CPU work that alternates between model inference steps and orchestration logic, with strict latency budgets for multi-turn interactions. AWS argues that Graviton5's 33% lower inter-core latency, 5× larger cache, and high core count per instance make it well suited for these workloads when they need to run at production scale without GPU economics .

Nitro Isolation Engine: The First Formally Verified Cloud Hypervisor

Beyond raw performance, the most technically significant addition to the Graviton5 platform is the Nitro Isolation Engine, a new component of the sixth-generation AWS Nitro System .

Implemented in Rust, the Nitro Isolation Engine is a minimal, purpose-built hypervisor component responsible for enforcing isolation between co-tenanted virtual machines . What distinguishes it from every other production hypervisor is formal verification: AWS has produced machine-checkable proofs using the Isabelle proof assistant that mathematically demonstrate :

• Memory safety and absence of runtime errors
• Functional correctness (the implementation matches its specification)
• Information flow control (no leakage between VM states)

In practical terms, this means AWS can provide mathematical certainty that one customer's workloads cannot access another's data or interfere with their execution, and that AWS operators are subject to the same isolation boundaries . AWS has committed to making the Nitro Isolation Engine's implementation and corresponding proofs available for customer review .

The engine is enabled by default on M9g instances . This represents a shift in cloud security assurance: from operational controls and audit narratives toward machine-checkable guarantees about the foundational isolation layer.

Early Adopters and Ecosystem

Named early adopters and benchmark partners include Meta, Snowflake, Uber, Honeycomb, SAP, Atlassian, and ClickHouse, along with HubSpot and others identified through performance data disclosures .

Customer-reported results span multiple workload categories:

• Databases: Snowflake reported 20–60% performance gains on Graviton4 and is expanding onto Graviton5 ; HubSpot saw MySQL query durations drop by up to 60% on M9g
• Observability: Honeycomb achieved 36% better throughput per core compared to Graviton4
• Analytics: ClickHouse observed 25–30% average query improvement on Graviton4 and 36% on Graviton5
• Application serving: Atlassian's migration of Jira and Confluence to Graviton4 yielded a 30% increase in throughput and a 25% reduction in costs

These results reflect patterns visible across the Graviton adoption curve: most workloads see immediate performance improvements with zero or minimal code changes when migrating from x86 to Arm, and the gains compound across generations as the silicon improves .

What This Means in Context

Graviton5 arrives at a moment when Arm-based server silicon has moved from a cost-optimization alternative to a mainstream performance choice. More than half of new AWS CPU capacity has run on Graviton for the past three years, and 98% of the top 1,000 EC2 customers already use Graviton-based instances .

With a monolithic 192-core die on a 3nm process, PCIe Gen 6 support, DDR5-8800 memory, and the addition of formally verified workload isolation, Graviton5 raises the ceiling not just for AWS's own instance families but for what customers can reasonably expect from cloud-native compute: performance, energy efficiency, and security guarantees backed by mathematical proof rather than operational promises.

The general availability of M9g and M9gd instances means these capabilities are now accessible through standard EC2 adoption paths, with compute-optimized C9g and memory-optimized R9g variants expected to follow .