{"slug": "aws-graviton5-m9g-is-ga-benchmarks-confirm-36-gains", "title": "AWS Graviton5 M9g Is GA: Benchmarks Confirm 36% Gains", "summary": "AWS Graviton5-powered M9g and M9gd EC2 instances are generally available, with production benchmarks from Honeycomb, ClickHouse, and HubSpot showing consistent 36% performance gains over M8g. The new chiplet design on TSMC's 3nm process doubles core count and increases cache fivefold, enabling significant price-performance improvements for most workloads.", "body_md": "AWS Graviton5-powered M9g and M9gd EC2 instances went generally available on June 10. Three companies have now published production data — and the numbers land consistently around 36%. That is not an AWS marketing figure. That is what Honeycomb, ClickHouse, and HubSpot measured in their actual workloads.\n\n## What the Benchmarks Actually Show\n\nStart here, because this is the part that matters. AWS claims up to 25% better compute performance, 35% faster web apps, 35% faster ML inference, and 30% faster databases compared to M8g. Those are headline numbers, and headline numbers can be optimistic. The production data tells a more interesting story.\n\n**ClickHouse** reported a **36% performance boost** over M8g with zero code changes, plus 16% higher query concurrency. **Honeycomb** ran a six-month A/B test of production observability ingest workloads and measured **36% better throughput per core**, 28% lower latency, and 26% less CPU utilization. **HubSpot** moved its MySQL databases to M9g and watched query duration drop by up to **60%**.\n\nTwo independent companies arriving at 36% is not a coincidence. It is a repeatable result from a hardware generation that actually changed something meaningful. [InfoQ’s detailed GA coverage](https://www.infoq.com/news/2026/06/aws-graviton5-ga/) breaks down each company’s methodology.\n\n## What Changed in the Silicon\n\nGraviton5 is a chiplet design — four chiplets on TSMC’s 3nm process — built around the ARM Neoverse V3 core running at 3.3 GHz (up from Neoverse V2 at 2.8 GHz in Graviton4). Core count doubled, from 96 to 192. But the change that explains the database and inference gains is the cache: L3 is now roughly 180 MB, five times larger than Graviton4, with each core getting 2.6x more L3 cache access than before.\n\nMemory bandwidth also jumped — DDR5-7200 delivers 691 GB/s, a 28% increase over Graviton4. PCIe Gen 6 brings I/O throughput to approximately 0.5 TB/s. Network bandwidth is up 15%, EBS IOPS up 30%. The whole stack moved.\n\n## The Price Math\n\nM9g instances are priced 9% above M8g across all sizes. Run that against a 25% compute improvement and you get roughly 15% better price-performance at steady state. That is the floor. For workloads hitting the database or inference profiles that HubSpot and ClickHouse tested, the improvement is high enough to run fewer instances — which means the upgrade can reduce costs, not just improve throughput.\n\nThe only case to pause is if your workload is bottlenecked somewhere Graviton5 does not address: single-threaded legacy serial code, or workloads that require x86 ABI compatibility. If neither applies, the upgrade makes financial sense now.\n\n## Formally Verified VM Isolation\n\nGraviton5 ships with the **Nitro Isolation Engine** — what AWS describes as the first formally verified hypervisor isolation in a major public cloud. The verification consists of 330,000 lines of machine-checked proofs in Isabelle/HOL, implemented in Rust, covering confidentiality, integrity, and memory safety between VMs. [Amazon Science published the full technical breakdown](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation). If you operate multi-tenant SaaS or work under compliance frameworks that require provable isolation, this belongs in your architecture documentation.\n\n## How to Migrate\n\nFor most workloads running Java, Python, Node.js, Go, Rust, .NET Core, or Ruby: change the instance type. No code changes required.\n\nFor containerized workloads: rebuild your container images for ARM64 and redeploy. AWS maintains a [Graviton Getting Started guide on GitHub](https://github.com/aws/aws-graviton-getting-started) with specific guidance for PostgreSQL, MySQL, ML frameworks, and more.\n\nThe memory-optimized R9g and compute-optimized C9g variants are not yet available — both are planned for later in 2026. If your workload fits general-purpose M9g sizing, you can move now. If you need the specialized families, watch for announcements in Q3.\n\n## Bottom Line\n\nM9g is the clearest EC2 upgrade recommendation in years. The benchmarks are consistent across independent companies, the price premium is modest, and the migration path is low-friction for most stacks. Start with a shadow deployment on one workload, measure for a week, and you will have the numbers to justify a full migration. [The official AWS launch post](https://aws.amazon.com/blogs/aws/now-available-amazon-ec2-m9g-and-m9gd-instances-powered-by-new-aws-graviton5-processors/) has region availability and instance size options.", "url": "https://wpnews.pro/news/aws-graviton5-m9g-is-ga-benchmarks-confirm-36-gains", "canonical_source": "https://byteiota.com/aws-graviton5-m9g-ec2-ga-benchmarks/", "published_at": "2026-07-09 00:08:10+00:00", "updated_at": "2026-07-09 00:20:23.339731+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-infrastructure", "ai-chips"], "entities": ["AWS", "Graviton5", "M9g", "Honeycomb", "ClickHouse", "HubSpot", "TSMC", "ARM"], "alternates": {"html": "https://wpnews.pro/news/aws-graviton5-m9g-is-ga-benchmarks-confirm-36-gains", "markdown": "https://wpnews.pro/news/aws-graviton5-m9g-is-ga-benchmarks-confirm-36-gains.md", "text": "https://wpnews.pro/news/aws-graviton5-m9g-is-ga-benchmarks-confirm-36-gains.txt", "jsonld": "https://wpnews.pro/news/aws-graviton5-m9g-is-ga-benchmarks-confirm-36-gains.jsonld"}}