Trusted Hardware for AI Workloads: Extending Confidential...
summary
Argues that confidential computing on CPUs already works but scaling to multi‑GPU AI clusters remains the bottleneck: single‑GPU solutions exist (closed‑source Nvidia CC) using encrypted CPU–GPU data transfers, a secure channel/key‑agreement and firmware signing, plus GPU memory protection, while multi‑GPU requires secure GPU‑GPU links that could impose large cryptographic overheads. Calls for concrete work on secure attestation for GPUs, scalable key management and secure interconnects, open validation frameworks and an open reference like the Certifier project.
AI-written summary · about
chapters (13)
- 0:00Trusted hardware for AI
- 2:00Speaker background
- 3:26Why confidential computing matters
- 5:36Hardware security needs
- 8:01Demand for trusted hardware
- 9:06Why confidential computing leads
- 10:09Single GPU confidential computing
- 11:35Multi-GPU confidential computing challenges
- 12:39Nvidia’s transparency gaps
- 13:23Industry adoption needs
- 15:55Lack of standards
- 16:58Open source in confidential AI
- 18:01Call to action
related talks
20mBuilding Trusted AI Factories with NVIDIA Accelerated Confidential Computing
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
44mGPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud
36mCTO Panel: The Impact of Confidential Computing with Intel, Azure, AMD & NVIDIA- 42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
37mIt's All About Trust: Why Confidential Computing Matters
19mHow Can We Have Private Access to Intelligence?