Microsoft reports up to 4.15x faster EDA workloads on Cobalt 200

Cobalt 200 targets demanding chip design workloads

Every Azure chip Microsoft has built began with electronic design automation, or EDA, jobs used to design, test and validate silicon before manufacturing. These jobs include simulation sweeps that can run for days, verification regressions that spread across thousands of cores overnight, and physical signoff runs that can use hundreds of gigabytes of memory for a single block.

Microsoft says these workloads account for hundreds of millions of compute jobs each year. Its latest evaluation used Azure Cobalt 200, the company’s custom compute infrastructure built on the Arm Neoverse CSS V3. The tested tools included Cadence Xcelium and Spectre, Synopsys PrimeTime and VCS, and Siemens Calibre.

Tools enabled on Cobalt 100 ran unmodified in the Cobalt 200 evaluation, allowing Microsoft to focus on performance testing rather than functionality.

Tests show gains across simulation and verification

Microsoft compared Cobalt 200 with Cobalt 100 and selected x86 Azure virtual machines. The virtual machines used the maximum possible core count for high throughput, while multiple jobs ran concurrently to test performance under high load. Throughput was normalized per logical core.

• Cadence Xcelium: More than 40% higher peak throughput per logical core than any other tested system.
• Synopsys VCS: More than 40% higher peak throughput per logical core than any other tested system.
• Cadence Spectre: 38% higher throughput per logical core than Cobalt 100, and 11% to 52% higher throughput than the other tested platforms.
• Siemens Calibre: Approximately 23% lower wall time than Cobalt 100, equivalent to 1.30x relative performance.
• Synopsys PrimeTime: 2.0x the peak throughput per logical core of Cobalt 100.

The workloads represent different parts of the chip development flow. Xcelium and VCS measure digital simulation, Spectre covers analog and mixed-signal simulation, Calibre performs physical design-rule checks, and PrimeTime verifies that a design meets timing requirements.

Architecture and cloud capacity support the results

Microsoft says EDA workloads have varied requirements. Simulation can depend on parallelism and memory behavior, verification requires throughput from many independent jobs, and implementation and signoff workloads are memory-intensive and bandwidth-sensitive.

Cobalt 200 combines physical cores without hyperthreading overhead, dedicated 3MB L2 cache per core and high provisioned memory bandwidth. According to Microsoft, these features are intended to maintain strong per-core performance when systems are heavily utilized, which matters to CAD managers planning overnight regressions and verification farms.

The company also points to platform-level benefits from running EDA in Azure. Teams can provision capacity for regression campaigns or signoff activity and release it when demand falls, rather than sizing fixed infrastructure for annual peak demand. Microsoft also cites high-performance storage, global capacity and adjacency to Azure AI services as parts of the platform.

Microsoft says the proximity between AI inference services and Cobalt 200 could support tighter integration of AI-assisted design workflows with EDA execution. The source describes this as an opportunity to reduce data movement, accelerate iterative design and verification cycles, and help AI agents coordinate engineering workflows.