QEMU and gem5

This chapter introduces the two simulators of the co-simulation: QEMU provides the guest runtime and functional device models, and gem5 provides performance simulation of the MI300X.

Functional and Performance Simulation

As the vhost, QEMU runs guest Linux, the amdgpu driver and ROCm applications and provides the CPU, memory and PCIe environment. It can also act as a device model process that exports a PCI device over vfio-user, guaranteeing functional correctness without microarchitectural timing. gem5 models the MI300X command processors, execution units, caches and memory hierarchy and provides the timing behavior relevant to performance.

Simulator	Role	Focus
QEMU (vhost)	Guest runtime	CPU, memory, PCIe, boot, QMP and serial
QEMU (device model)	Functional PCI device model	Configuration space, BARs and interrupts
gem5	MI300X performance model	PM4/SDMA, IH, GPU VM, shader/CU arrays and the Ruby memory hierarchy

Choosing a Device Implementation

A topology chooses the device implementation through its device adapter: gem5-mi300x uses the gem5 MI300X performance model, qemu-device uses a QEMU functional device model, and the mixed-models machine uses both. Chapter 5 records the adapters of every machine, and the adapter contract lives in Device adapters.

Chapter Contents

• QEMU roles
• gem5 performance simulation
• Guest driver requirements
• Model debugging

Recommended Reading Path

• Start with this page and QEMU roles and gem5 performance simulation for the division of labor.

• For guest startup, read QEMU as vhost.

• For QEMU functional device models, read QEMU device models.

• For the MI300X model, continue with MI300X PCI interface and the pages after it.

• For driver and model debugging, read Guest driver requirements and Model debugging.