Gpu virtualisation

A method of virtualizing a graphics processor unit (GPU) for operation with a central processing unit (CPU) running a plurality of virtual machines under the control of a hypervisor that has allocated an identifier to each of the plurality of virtual machines, the method comprising:

• configuring a translation table in a root memory management unit associated with the GPU, wherein the translation table defines translations between guest physical addresses and actual physical memory addresses based on an identifier for the virtual machine;
• and
• during execution of one or more of the virtual machines, receiving transactions output by the CPU with the identifier for the virtual machine to which the transaction relates and, within the root memory management unit, translating addresses in a transaction based on the identifier with which the transaction is tagged, wherein each transaction is a transaction on a system bus which interconnects the CPU, GPU and other peripherals;
• and
• using the assigned identifier in combination with the root memory management unit to enable mediation of some or all transactions without hypervisor intervention.

A non-transitory computer readable storage medium having stored thereon a computer readable dataset description of an integrated circuit that, when processed in an integrated circuit manufacturing system, causes the integrated circuit manufacturing system to manufacture a graphics processor unit (GPU) arranged to operate with a central processing unit (CPU) running a plurality of virtual machines under the control of a hypervisor, wherein transactions output from the CPU are assigned an identifier for the virtual machine to which it relates, the GPU comprising:

• a GPU memory management unit arranged to translate between virtual memory addresses used by the graphics processor unit and guest physical addresses;
• and
• a root memory management unit configured by the hypervisor and arranged to translate between guest physical addresses used by the GPU memory management unit and actual physical memory addresses based on the identifier for the virtual machine assigned to the transaction;
• wherein access to one or more portions of a memory resource is controlled as a consequence of the address translations performed by the root memory management units;
• wherein each transaction is a transaction on a system bus which interconnects the CPU, GPU and other peripherals;
• and
• wherein the assigned identifier in combination with the root memory management unit enables mediation of some or all transactions without hypervisor intervention.