Scaling GPUs

I will post more details, and this is a placeholder for that, but BIOS IT has a product which allows two useful things:

  1. Access from a container or VM to a hosted GPGPU. This would allow, for example, a workload using Robot Operation System (ROS) which is easiest to install on Ubuntu to be run on a server that is natively CentOS. In theory that sort of server-side stuff might not technically need ROS, but the higher level might assume the installation of ROS elements to bring in things like OpenCV with GPGPU acceleration so it would still mean that a researcher who has a workload that uses this could use a container or VM with unmodified code and use the GPGPU effectively without messy things like passing through the GPGPU PCIe over from the host to the VM, etc.
  2. The ability to segment a large GPGPU (presumably smaller ones too, although that’s less of a likely use case) so that multiple users can access it. This would be of interest in terms of allowing the purchase of a V100 to serve the need for a large GPGPU without compromising the ability to support smaller jobs and without the need to have an HTC GPGPU cluster with prosumer cards.