With the Hosted AI platform there are multiple methods of accessing GPU resources ranging from direct mapping of dedicated GPUs to a virtual machine, through to subscribing to a shared pool. We describe this below.
GPUaaS passthrough
GPUaaS passthrough allows a virtual machine to directly access a physical GPU. This means that the GPU resources are dedicated to a single virtual machine, offering improved performance and efficiency for applications that require intensive processing. The VM exclusively sees and controls the GPU device as a direct PCIe device.
GPUaaS Pool
GPUaaS pools provide the administrator with the ability to create a group of GPU resources that can be allocated and managed together. This leverages a container framework to map end user workloads to a virtual GPU resource that is in fact a share of one or more physical GPU cards. The administrator has control over which physical cards are added to any pool of resource, how many times the GPU resource is to be overcommitted and other tunable parameters to control how the end user tasks get scheduled across the GPUs.