AWS Graviton5 M9g Instances: 25% Better Performance and What the MySQL Numbers Mean
Cloud Platform EngineeringHubSpot cut MySQL query duration by 60% just by migrating to the new instance type. ClickHouse saw 36% better throughput per core. Honeycomb matched that number. These are not synthetic benchmarks — they are production workloads, and they tell you more about whether to migrate than any spec sheet.
What Happened
Amazon’s EC2 M9g and M9gd instances, powered by the new Graviton5 processors, are now generally available in four regions: US East (N. Virginia and Ohio), US West (Oregon), and Europe (Frankfurt). The M9g class covers pure compute across sizes from 1 to 192 vCPUs; the M9gd variant adds local NVMe storage up to 11.4 TB with 30% higher IOPS than the previous M8gd counterparts.
Graviton5’s architecture improvements over Graviton4 are substantial: 192 cores, a 5× larger L3 cache, up to 33% lower inter-core latency, DDR5-8800 memory support, and PCIe Gen6 — the first AWS CPU to support that last specification. Network bandwidth reaches 100 Gbps at the top end; EBS bandwidth peaks at 72 Gbps.
AWS claims 25% better overall compute performance compared to Graviton4, with specific acceleration for web applications and ML inference (35% faster) and database workloads (30% faster).
Why It Matters
The real-world numbers above are worth unpacking. MySQL dropping 60% in query duration at HubSpot is not a minor optimization — that is infrastructure-grade latency reduction that changes caching strategy, read replica topology, and headroom planning. The ClickHouse and Honeycomb figures (both 36% better throughput per core) are consistent enough to be reliable for analytics and observability workloads.
The architectural story behind those numbers is the L3 cache expansion. Inter-core latency dropping 33% directly benefits multi-threaded database query execution and anything that shares state across cores. For containerized microservices, smaller L3 accesses per transaction mean more requests handled before spilling to memory — which is the primary lever on query latency at scale.
M9gd with local NVMe is the variant to watch for workloads that currently use io2 volumes primarily for scratch space. The 11.4 TB NVMe option with 30% higher IOPS than M8gd offers a cost-effective path for ephemeral analytics, shuffle space in distributed jobs, and read-heavy caching tiers that don’t need EBS’s durability guarantees.
For teams already running Graviton4 on M8g instances, the migration path is straightforward — same ARM instruction set, same toolchain. The break-even on migration cost comes quickly when the MySQL or analytics benchmarks hold in your workload profile. AWS’s formal verification work on the Nitro isolation engine means the isolation guarantees underneath M9g are now mathematically proven, not just tested.
The So What
If your database or analytics workloads are on M8g or earlier Graviton instances, the benchmark consistency across three independent production deployments makes this worth testing now. Run a shadow load on a single M9g instance and measure against your specific query mix — the 25–60% range suggests the improvement is workload-dependent, and your number is the only one that matters. The four-region launch limits availability today, but the production benchmark data is early enough to plan Q3 migrations against real capacity signals.
Source: AWS Blog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →