Cloud Services Infrastructure as a Service Clouds (IaaS)

Microsoft reports EDA performance gains on Azure Cobalt 200

Microsoft says early testing of electronic design automation (EDA) workloads on Azure Cobalt 200 delivered performance improvements over its previous-generation Cobalt 100 processors. Across evaluated scenarios spanning digital and analog simulation, physical verification and static timing analysis, the company reported gains ranging from 1.30 times to 4.15 times, depending on the workload and measure.

The results come from Microsoft’s own silicon-design environment and internal test cases. The company cautions that findings apply to the workloads, software versions, virtual machine configurations, concurrency levels and test conditions evaluated; they should not be read as a universal comparison for every EDA deployment. Microsoft said it is expanding workload coverage and validating results at greater scale as Cobalt 200 approaches general availability.

EDA tools are used to design, simulate and verify chips before manufacturing. Microsoft describes its silicon engineering work as involving simulation jobs that can run for days, overnight verification regressions distributed across many cores, and memory-intensive physical signoff tasks. The company says these workloads account for hundreds of millions of compute jobs annually.

Results across EDA workloads

Microsoft tested Cadence Xcelium and Spectre, Synopsys VCS and PrimeTime, and Siemens Calibre. The tools already enabled for Cobalt 100 ran unmodified in the Cobalt 200 evaluation, allowing the company to concentrate on performance rather than changes to tool functionality. Tested systems included Cobalt 100 with 96 cores, Cobalt 200 with 128 cores, the hx176rs with 176 cores, and the e160adsv7 with 160 logical cores and hyperthreading enabled.

  • For internal RTL digital-simulation tests using Cadence Xcelium and Synopsys VCS, Cobalt 200 delivered more than 40% higher peak throughput per logical core than any other system tested.

  • For a Cadence Spectre analog and mixed-signal simulation workload, Microsoft measured 38% greater throughput per logical core than on Cobalt 100. It also reported a 11% to 52% advantage over the other tested platforms for that workload.

  • For Siemens Calibre design-rule checking on an IP macro, using 16 cores per job, Cobalt 200 reduced wall time by about 23% versus Cobalt 100, equivalent to 1.30 times relative performance.

  • For a Synopsys PrimeTime static-timing workload using 16 cores per job, Cobalt 200 reached twice Cobalt 100’s peak throughput per logical core and exceeded the selected x86 Azure virtual machines.

The results differ by tool and test method. For example, Microsoft measured Calibre using mean total wall time, while several simulation and timing results were expressed as jobs per hour per logical core. The company’s broader summary cites a range of 1.30 to 4.15 times across evaluated scenarios, but the detailed figures above describe selected test cases rather than a single benchmark applied uniformly to all systems.

Architecture and cloud capacity

Microsoft attributes the gains in part to Cobalt 200’s design, which is based on Arm Neoverse CSS V3. The company points to physical cores without hyperthreading overhead, 3 MB of L2 cache per core and provisioned memory bandwidth. Those characteristics are intended to support per-core performance when multiple jobs run concurrently, a condition that matters for verification farms and other heavily loaded EDA environments.

EDA work spans different resource patterns: simulation can benefit from parallel execution and memory behavior, verification often runs many independent jobs, and implementation or signoff can demand substantial memory bandwidth over long periods. Microsoft says its tests used VMs configured with the maximum number of cores available for high throughput and ran multiple jobs concurrently to assess performance under load. Throughput was normalized per logical core to account for differing VM sizes.

The company also emphasizes that chip-design infrastructure involves more than processor speed. Engineering teams may need to scale capacity for regression campaigns or signoff pushes, then release resources as demand falls after project milestones. Azure’s cloud model can provide elastic compute, alongside storage for design databases, waveform dumps and verification data. Microsoft additionally cites regional capacity and proximity to Azure AI services as platform considerations, while describing tighter integration with AI-assisted design workflows as an opportunity rather than a measured result in these tests.

Further evaluation underway

Microsoft said Spectre and Calibre, which had performed well on Cobalt 100, also improved between generations, while Xcelium and PrimeTime showed particularly large gains in the evaluated workloads. It is working with EDA partners on additional tuning and plans to broaden the tests. The announcement describes early findings, not a comprehensive independent benchmark or a guarantee of equivalent performance for customer designs.

For semiconductor teams considering cloud infrastructure, the practical next step is to test representative design flows against their own requirements, including concurrency, memory needs and data movement. Microsoft is inviting customers to evaluate their EDA workloads on Azure Cobalt 200 virtual machines as the platform moves toward general availability.