upvote
Newer NVidia cards (H100 and up) support both in-memory model encryption and ‘trusted’ execution environment/remote attestation, not sure how widely used in frontier model deployments, but at least vendor claimed perf overhead is ‘3%’ [0]

[0] https://www.spheron.network/blog/confidential-gpu-computing-...

reply