Microsoft Azure Cobalt 100 VMs powered by Arm Neoverse deliver higher performance and efficiency for key workloads

Following Microsoft’s introduction of Azure Cobalt 100 virtual machines (VMs) powered by Arm, customers have adopted new instances  across a variety of customers and use cases. Now available in 32 regions globally and growing rapidly, Cobalt 100 VMs provide a major step forward in performance, efficiency, and scalability for next-generation data center workloads. These gains are rooted in the innovation of the Arm Neoverse Compute Subsystems (CSS), purpose-built to power modern cloud-native applications with higher throughput and lower energy use. 

Leading software vendors, such as One Trust and Databricks, are among the customers that have reported significant performance efficiency improvements, along with cost savings. This aligns with a growing industry shift toward heterogeneous compute architectures, where Arm’s scalable, energy-efficient compute platform is enabling hyperscalers to meet rising performance demands without compromising on sustainability goals. Meanwhile, Microsoft services such as Teams and Microsoft Defender Endpoint have seen up to 45% better performance on these instances.  

Arm recently benchmarked several workloads, such as in-memory databases, web servers and quantitative finance, on Cobalt 100 instances and compared them against AMD Genoa-based Azure instances. The results are illustrative of the performance and efficiency uplifts that can be achieved by running on Arm-based Azure Cobalt 100 VMs, confirming what Cobalt 100 is designed to achieve: significant price-performance benefits, broad workload versatility, and substantial cost savings. 

The improved performance of general-purpose workloads running on Cobalt 100 

Redis remains a cornerstone of modern distributed applications, especially for real-time data processing and caching. The performance and price/performance advantages on the D4ps_v6 Cobalt 100 instances are impressive, with 48% better performance and 91% better price-performance when compared to AMD Genoa D4as_v6 instances. This underscores Arm’s ability to drive real-world performance where microservices and latency sensitive real-time data interactions dominate cloud-native application stacks. 

Meanwhile, measuring web infrastructure and networking workloads, such as Nginx, showcases Cobalt 100’s balanced performance across compute, memory, and I/O. In the benchmark tests, Arm powered D4ps_v6 Cobalt 100 instances showed 53% better performance and 99% better price- performance for load balancing requests, compared to AMD Genoa D4as_v6 instances. 

Finally, in GCC/G++-15 benchmarking of QuantLib, an open-source quantitative finance library used for modeling, pricing, and risk management of financial instruments, showed 47% better performance and 89% better price-performance on D4ps_v6 Cobalt 100 instances, compared to AMD Genoa D4as_v6 instances. 

Efficiency, scalability, and real-world value 

Arm continues to work closely with Microsoft Azure on innovations like the new Cobalt 200 processor that demonstrate how custom silicon built on Arm unlocks new levels of performance and efficiency for AI-driven cloud infrastructure at scale.

The Cobalt 100 instances maintain Arm’s hallmark power efficiency, meaning any performance gains come without proportionally higher energy use. The result is a better price-performance profile for cloud operators and customers alike. For example, Arm optimizations in ONNX Runtime combined with Cobalt 100 processors expand AI performance for large language model (LLM) inference, leading to up to 1.9x higher performance and 2.8x better price/performance compared to AMD Genoa-based instances. 

Whether customers are running high-frequency trading systems, caching layers, API front-ends, or Java microservices, Cobalt 100 delivers the performance and efficiency to scale confidently into the future.  As Arm’s ecosystem expands across software, tooling, and silicon partners, developers benefit from a platform built for long-term cloud innovation and sustainability. 

Arm at Microsoft Ignite 

Arm will be at Microsoft Ignite in San Francisco (booth 5252) showcasing the superior price-performance of Cobalt 100 VMs, as well as the new Arm MCP server in action through live demos. 

We also have the following resources to help you get started on your journey to Cobalt 100:   

• Migration to Cobalt 100 with Arm Azure specific Learning Paths: Streamline your transition to Cobalt 100 instances with comprehensive guides and best practices.   

• Arm Software Ecosystem Dashboard: Stay up-to-date on the latest software support for Arm. 

• Arm Developer Hub: Whether you are just getting started on Arm or looking for resources to help you create top-performing software solutions, Arm Developer Hub has everything you need to start building better software and deliver rich experiences for billions of devices. Download, learn, connect, or ask questions within our growing global developer community. 

Footnote

All benchmark results were obtained under controlled and repeatable test conditions using the latest available software and platform configurations.

Tests were conducted on Ubuntu 25.04 with consistent kernel, compiler, and runtime environments across all compared platforms (Cobalt100, Genoa, and Emerald Rapids). Each workload was executed multiple times to ensure measurement consistency and statistical validity.

General testing principles

Benchmarks were run with identical thread counts, client configurations, and runtime parameters across all platforms.

NUMA and memory settings were aligned per instance type; HugePages, THP, and swapiness settings were standardized.

All systems used the same OS image, compiler version (GCC/G++ 15.0.1), and library versions (OpenSSL 3.4.1, GlibC 2.41).

Benchmark scope

Redis, PostgreSQL, MySQL: Throughput and throughput-per-dollar measured with Memtier and HammerDB (TPROC-C).

SPECjbb 2015: Composite, max-jOPS, and critical-jOPS /$ evaluated using OpenJDK 21 on Ubuntu 25.04.

Nginx: HTTPS load-balancer throughput at 99th percentile latency using wrk with fixed connection and thread counts.

QuantLib: Financial analytics workloads built with GCC 15 and vendor LLVM toolchains, measured for both single-thread and all-vCPU performance.