On July 13, Phoronix published a detailed benchmark comparison of Amazon EC2's latest general-purpose instances, and the results deliver a clear message: AMD's EPYC Turin processor still holds the performance crown in the cloud. AWS's own Graviton5 Arm chip managed to outpace Intel's Xeon 6 Granite Rapids in several tests, but it remains a Linux-only proposition that Windows shops can only watch from the sidelines.

What Actually Changed

Phoronix ran a suite of Linux workloads across four EC2 instance types, each configured with 64GB of memory and the \"4xlarge\" size: the new Graviton5-powered m9g.4xlarge, the AMD EPYC Turin-based m8a.4xlarge, the Intel Xeon 6 Granite Rapids m8i.4xlarge, and the previous-generation Graviton4 m8g.4xlarge. All instances expose 16 vCPUs to the operating system, but the silicon underneath tells a different story. The AMD m8a and Graviton5 m9g each pack 16 physical cores, while the Intel m8i relies on 8 physical Granite Rapids cores with Hyper-Threading to reach the same vCPU count.

The test environment was Ubuntu 26.04 LTS with Linux 7.0 and GCC 15.2, meaning all results reflect a modern Linux software stack. AWS made the M9g family generally available on June 10, 2026, touting up to a 25 percent compute performance improvement over Graviton4. The leap comes courtesy of Arm's Neoverse-V3 core design and faster DDR5-8800 memory. Phoronix's numbers bear out a substantial generational jump: Graviton5 opened a clear gap to Graviton4 and, more notably, edged past the Intel Xeon 6 configuration in several key benchmarks.

For example, in the QuantLib quantitative finance library test, the m9g.4xlarge finished ahead of the m8i.4xlarge, though AMD's m8a.4xlarge still led the pack. The in-memory CockroachDB benchmark told a similar story, with Graviton5's superior memory bandwidth translating into a real advantage over the Intel instance. But across the entire workload set, AMD's EPYC Turin M8a instance was the broad performance leader. In multiple scenarios, it also delivered the best performance per dollar, even though its hourly price is higher than the Intel option.

What It Means for You

For Windows Server Administrators

If you run Windows Server on EC2, the immediate takeaway is simple: the Graviton5 party is not for you. AWS lists the M9g family as Linux-only, because Windows Server lacks first-party Arm64 support on the platform. You cannot lift a Windows workload and drop it onto an m9g instance — not without a complete re-architecture of your application stack onto Arm-compatible binaries and operating systems, a path that Microsoft does not currently offer on EC2.

So the real choice remains between the x86-based M8a (AMD) and M8i (Intel) families. Given Phoronix's data, the M8a with EPYC Turin looks like the strongest all-rounder for CPU-bound Windows services such as .NET applications, SQL Server, and custom business logic. The M8i's eight physical Granite Rapids cores, while capable, are outgunned by the 16 real cores in the AMD instance. For workloads that thrive on high single-threaded performance or are optimized for Intel libraries, the M8i still has a place, but the overall picture favors AMD for most general-purpose Windows deployments.

For DevOps and Mixed-Environment Teams

Organizations that run both Windows and Linux on EC2 gain a more nuanced set of options. Graviton5 emerges as a credible, often faster, and potentially cheaper alternative to Intel Xeon 6 for Linux services. Containers running on Amazon ECS or EKS, Java-based microservices, in-memory databases, and CI/CD pipelines are prime candidates for migration to Arm64 if they haven't been moved already. The performance uplift over the previous Graviton4 and the competitive showing against Intel mean that shifting non-Windows workloads to m9g instances could free up budget or capacity for your x86-licensed Windows estate.

Still, this requires careful workload validation. Not every application or database is Arm-ready out of the box. Plan for a test cycle, and don't assume that a Graviton5 win in, say, CockroachDB automatically applies to your specific PostgreSQL or MariaDB deployment.

For Decision-Makers Evaluating Cloud Spend

Performance per dollar is often more important than raw speed, especially when scaling large fleets. Phoronix noted that in some workloads, the higher-priced M8a ended up being the better value because it completed tasks so much faster. This counterintuitive result underscores the need to run your own cost-performance models. Also factor in the potential licensing implications: Windows Server licensing on AWS is tied to vCPU counts, and both the M8a and M8i expose the same 16 vCPUs. However, the physical core count disparity may affect per-core licensing costs for software that bills per physical core rather than per vCPU — a detail worth checking with your vendors.

How We Got Here

AWS's Graviton journey started in 2018 with the first-generation A1 instances, using Annapurna Labs-designed chips. Each subsequent generation has narrowed the gap with x86 processors. Graviton4, launched in 2024, brought the Neoverse-V2 core and was already a strong performer against contemporary Intel and AMD offerings. The move to Neoverse-V3 with Graviton5 accelerates this trend, and the chip now directly challenges Intel's server dominance in the cloud.

On the AMD side, the EPYC Turin series (5th Gen EPYC) began rolling out in late 2025, leveraging the Zen 5 microarchitecture. AMD's strategy of delivering high core counts per socket — and, crucially, one thread per core without Hyper-Threading — provides a clean performance profile for many cloud workloads. Intel's Granite Rapids, meanwhile, represents the company's attempt to regain ground after several generations of manufacturing delays. It brings improved performance-per-watt and modern I/O features, but in AWS's current instance designs, it arrives with only half the physical cores of its rivals in the tested 4xlarge size.

The disparity matters. AWS's decision to field 16-core AMD and Arm instances while limiting the Intel option to 8 cores with Hyper-Threading is a product of platform design and power/thermal constraints, not a universal judgment on the silicon itself. For customers just looking for a 16-vCPU general-purpose instance, the benchmark numbers speak louder than architecture details.

On the software front, Ubuntu 26.04 and the Linux 7.0 kernel bring mature Arm64 support and latest compiler optimizations. Microsoft has been building Windows on Arm for client devices, and server Arm versions exist in previews, but as of July 2026 there is no production path for Windows Server on AWS Graviton. That may change — but not yet.

What to Do Now

Assess Your Workload Profile

First, classify your EC2 instances by operating system and application architecture. For pure Windows Server workloads, the immediate action is to evaluate a move from older instance families (like M5 or M6) to either M8a or M8i, based on your performance requirements. Benchmark your own applications, not just the Phoronix suite, because real-world behavior varies widely.

  • Peak throughput sensitive workloads (transactional databases, web front-ends, batch processing): Start testing with m8a.4xlarge. The 16 physical AMD cores generally provide higher aggregate throughput.
  • Single-threaded or latency-bound Windows services: Compare m8a and m8i head-to-head. The Intel instance's higher clock speeds on fewer cores might occasionally win.
  • Software that explicitly requires Intel features (e.g., certain ML libraries, in-memory database acceleration): The m8i remains your default choice, but validate whether the AMD instance can deliver sufficient performance at a different price.

Explore Graviton5 for Non-Windows Workloads

If you have Linux-based services, container orchestrators, or any Arm-compatible application stack, pilot the m9g.4xlarge. Start with non-critical, stateless services to verify compatibility. Use the AWS Graviton readiness tools and test against the same load profiles you use today on x86. Given the performance and price advantages shown in the Phoronix data, you could see meaningful savings while maintaining or improving speed.

Keep an Eye on Licensing and Reserved Instance Pricing

Performance per dollar doesn't stop at the instance price. If you use Microsoft SQL Server or other per-core licensed products, the physical-to-vCPU mismatch between AMD/ARM and Intel instances may alter your total cost. The AMD instance's 16 physical cores could trigger higher licensing fees than the Intel's 8 physical cores, even though both show 16 vCPUs. Consult your licensing agreements.

Plan an Arm-Ready Pipeline

Longer term, if Microsoft brings Windows Server to AWS Graviton, the landscape will shift. Start testing your .NET applications in Arm64 Linux containers today using the multiplatform build capabilities of .NET 8 and later. This prepares your code for a future Windows-on-Arm reality and gives you immediate savings on Linux services.

Outlook

The Phoronix results are a snapshot of a rapidly moving target. AWS regularly refreshes its instance types, Intel plans denser Granite Rapids instances with more cores, and AMD Zen 5 successors are already in development. The server Arm ecosystem is also gaining momentum; Qualcomm has teased cloud-focused chips, and Ampere continues to push Altra derivatives.

For Windows admins, the next milestone to watch is Microsoft's Build and AWS re:Invent conferences later this year. A joint announcement of Windows Server on Graviton would be a game-changer, but there are no public signals yet. In the meantime, AMD's EPYC Turin solidifies its position as the go-to performance choice on EC2. Graviton5's ascent adds welcome competitive pressure that benefits everyone through better performance and pricing — even if Windows users must settle for watching the race from the x86 grandstands.