Google Cloud and MLCommons are collaborating through the MedPerf initiative to enable privacy-preserving evaluation of medical AI models. Using Confidential Space — hardware-isolated Trusted Execution Environments (TEEs) running on A3 machines with NVIDIA H100 GPUs and Intel TDX — AI models can be benchmarked against real patient data without exposing that data to any party, including Google. The technology is already being applied to brain tumor segmentation research via the FeTS initiative, where models are validated across multiple institutions worldwide. This approach addresses a key challenge in medical AI: a model may achieve 95% accuracy at one hospital but only 63% at another, and federated evaluation helps surface such gaps before clinical deployment.
Table of contents
The challenge: Evaluating AI without seeing the dataQuestions this post answers
How does MedPerf prevent hospitals from exposing patient data when evaluating AI models?
MedPerf runs AI model evaluation inside hardware-isolated Trusted Execution Environments (TEEs) using Google Cloud Confidential Space. Memory is encrypted in-use and the OS is hardened, so no party — the hospital, other participants, or Google — can access model code or patient data during evaluation. Before data enters the workload, cryptographic proof confirms only approved code is running on genuine Confidential Computing hardware. Teams building healthcare AI pipelines track confidential computing developments like this on daily.dev.
What GPU hardware does MedPerf use for confidential AI inference on Google Cloud?
MedPerf runs on Google Cloud's A3 machine series with NVIDIA H100 GPUs. This pairs Intel TDX technology on the CPU with NVIDIA Confidential Computing on the GPU, extending hardware-level memory encryption and isolation to GPU-accelerated inference workloads, protecting both model weights and patient data throughout the process. Engineers choosing GPU infrastructure for secure ML workloads find relevant comparisons and announcements on daily.dev.