Skip to content

(Why) GPUs are awful (to virtualize)#

nog.fi meeting 2026.06 digest Program Adjacent pick 16–17 June 2026

Abstract Recording

Johannes Brummer

Play the recording: (Why) GPUs are awful (to virtualize)

  • Why: For anyone running networks for AI-era compute, knowing why GPUs are painful to virtualize (device assignment, passthrough, memory mapping) shapes what the fabric must support — SR-IOV/passthrough choices interact directly with RDMA fabrics.
  • What: Hands-on experience virtualizing different GPU types on multiple different hardware platforms, plus approaches to providing them to actually paying customers. (from abstract only)
  • How: Practical experience report; no methodology stated in the abstract.
  • Where applicable: Hosting/service providers offering GPU compute. Useful context for designing networks under GPU clusters, but it is a compute virtualization talk, not a networking one.