In 2022, NVIDIA® released the Hopper Architecture as the successor of the Ampere Architecture. One of the major features introduced by Hopper is confidential computing and the NVIDIA® H100 Tensor Core GPU is the first GPU of this architecture.
Building on our prior implementation of confidential computing with Intel® Trust Domain Extensions (Intel® TDX), we are excited to announce the latest tech preview release for Ubuntu Plucky 25.04 with shared device pass-through support for the H100. This new feature adds the ability to run confidential AI workloads on the H100 GPUs.
With this release, all AI workloads making use of the NVIDIA® H100 GPUs from inside a TD (TDX virtual machine) can now benefit from end-to-end confidentiality protection on CPU, GPU, and in transit.
Please visit our GitHub repository for more information: GitHub - canonical/tdx: Intel confidential computing - TDX