Before the First Incision: Shifting PCIe Virtual Function Validation Left with ZeBu

My son, who is a vascular surgeon, tells me his success with surgery begins long before a procedure starts. He says that surgeons spend countless hours studying imaging data, planning for complications, and validating their approach before entering the operating room. Their goal is to identify risks early to reduce surprises and improve outcomes.

Microsoft faced a similar challenge validating PCIe virtualization environments. The engineering team wanted to validate critical hardware and software interactions months earlier than first silicon to reduce risk before deployment.

As modern AI accelerators increasingly rely on virtualization, firmware, drivers, and operating system interactions, their verification environment must extend beyond hardware functionality to encompass the complete software stack.

Teams working on these complex AI accelerator programs increasingly need to bring up software, firmware, and virtualization stacks before first silicon. Waiting for hardware availability before full system validation begins can greatly extend debug cycles and delay software.

At SNUG 2026, James Mackenzie, Director of Design Verification at Microsoft, shared how his team used the Synopsys ZeBu Virtual Host Solution (VHS) to validate PCIe SR-IOV functionality, nested virtualization flows, and production software stack before tape-out. Microsoft’s approach to validation moved critical activities to earlier in the development cycle giving firmware developers access to realistic system environments months ahead of silicon availability.

A particularly valuable aspect of the approach was support for multiple levels of virtualization.

The validation environment they created contained:

• L0: ZeBu host and physical system
• L1: A virtual machine assigned physical functions (PFs)
• L2: Guest virtual machines assigned virtual functions (VFs)

This validation environment (see figure 2) created three layers of pass-through. Microsoft used QEMU and KVM-based virtualization to propagate physical and virtual functions through these layers, closely mirroring real deployment environments.

According to McKenzie, the complexity came from ensuring that PFs and VFs pass correctly through every level while maintaining expected functionality and accessibility. 

Microsoft used this environment to validate several key operations.

The team successfully:

• Enumerated virtual functions inside guest virtual machines
• Accessed PCIe configuration space
• Executed on-device DMA operations
• Validated MSI-X interrupt functionality
• Transferred data across the virtualization hierarchy and PCIe transactor chain
• Verified correct operation of production driver stacks

The real value came from enabling software teams to exercise production driver stacks and deployment workflows long before silicon arrived. The interface infrastructure became a foundation for earlier software readiness rather than simply a mechanism for protocol validation.

Traditional speed-adapter-based environments often require dedicated hardware systems and ongoing maintenance. McKenzie explained that VHS eliminated the need for real hosts and maintaining multiple speed-adapter systems. Teams could create numerous configurations leveraging their ZeBu capacity.

This scalability enabled Microsoft to move a large population of firmware developers onto a common validation platform, creating what McKenzie described as a significant shift-left impact. 

The most compelling outcome was not merely functional correctness. It was acceleration of software and firmware development.

"Having this SR-IOV functionality in Zebu really allowed us to bring up our entire production stack months before tape out."  James Mackenzie, Microsoft *

Microsoft reported that the solution enabled validation teams to gain confidence in their driver stacks and use cases weeks or even months earlier than would have been possible using traditional lab-based validation flows. Results generated in the pre-silicon stack translated directly to silicon bring-up; they could reproduce the same results with real silicon in hours.      

Microsoft reported that the solution enabled validation teams to gain confidence in their driver stacks and use cases weeks or even months earlier than would have been possible using traditional lab-based validation flows. Results generated in the pre-silicon stack translated directly to silicon bring-up; they could reproduce the same results with real silicon in hours.      

Like a vascular surgeon preparing for a complex procedure, Microsoft's team recognized that success depends on what happens long before execution with silicon begins. Surgeons rely on planning, simulation, and validation before entering the operating room to reduce uncertainty and improve outcomes. Microsoft applied the same principle to silicon development.

By using Synopsys ZeBu VHS, Microsoft extended verification beyond hardware functionality to software-defined system validation. The team brought up production software stacks months before tape-out, validated critical virtualization infrastructure, exercised production driver stacks, and gained confidence in real-world use cases before silicon arrived. These efforts enabled engineering teams to begin meaningful software validation earlier and approach silicon bring-up with substantially higher confidence.

References - Converge SJ (SNUG) 2026