Azure Cobalt 200 Outperforms Cobalt 100 Per Instance and Per Core
-
Jonathan Fellows
Signal65 testing across developer, agentic AI and HPC workloads shows Azure Cobalt 200 delivering 1.43x to 1.93x the throughput of Cobalt 100 per instance and 1.07x to 1.45x per vCPU, with up to 5.5x per instance in Microsoft’s own chip-design evaluation.
Azure Cobalt 200 is the second generation of Microsoft’s custom Arm-based server processor. By designing a custom CPU offering like Cobalt, Microsoft can tune it for the workloads Azure customers run and engineer it together with the servers, software and services the company offers around it. Customers are also asking general-purpose instances to carry more of their work, from software builds and data processing to AI agents and scientific simulation. The Cobalt 200 instance tested here carries 128 Arm Neoverse V3 cores with 3 MB of L2 cache per core, compared with 96 Neoverse N2 cores and 1 MB of L2 cache per core on Cobalt 100.
Signal65, an independent AI benchmarking and validation firm, ran both generations through developer and system tasks, function calling and molecular dynamics. We also reviewed chip-design results that Microsoft’s silicon engineering team produced on its own design flows.
Cobalt 200 led Cobalt 100 on all seven Signal65 results in this report, delivering between 1.43x and 1.93x the throughput per instance. The Cobalt 200 instance has 1.33x the vCPUs of its predecessor, so each Cobalt 200 vCPU also completed more work, by 1.07x to 1.45x. In Microsoft’s chip-design evaluation, a fully loaded Cobalt 200 instance completed up to 5.5x as many RTL simulation jobs per hour as a Cobalt 100 instance.
Key Highlights:
Key Findings
- Developer and system tasks: Cobalt 200 completed 1.76x the work per instance across the Terminal-Bench 2 task suite and ran each task 1.32x as fast per vCPU.
- Agentic AI: function-calling throughput rose 1.59x to 1.93x per instance (1.20x to 1.45x per vCPU) with identical accuracy.
- Scientific computing: GROMACS molecular dynamics rose 1.43x to 1.46x per instance (1.07x to 1.09x
per vCPU). - Chip design: in Microsoft’s own tests, reviewed by Signal65, a fully loaded Cobalt 200 instance completed 3.43x to 5.53x as many RTL logic simulation jobs per hour as a Cobalt 100 instance (2.58x to 4.15x per vCPU).
All Signal65 results compare the largest general-purpose VM of each generation, Standard_D128pds_v7 against Standard_D96pds_v6, at matched memory per vCPU. Cobalt 200 was tested in opportunistic mode, a deployment option that lets its cores scale up to 3.6 GHz. This report covers compute-intensive workloads that scale across the full instance, and Signal65 plans to extend the comparison to additional workloads in future testing.
Research commissioned by:


