I have the same question since 2010 (http://communities.vmware.com/thread/103067) and it seems to me that the experts (not so experts as can repair) always think of using GPGPU in guests (VMs) and not for an improvement of the performance of the host (hypervisor).