Reasonably Secure GPU Acceleration
summary
Proposal builds on vertdio GPU native contexts with a host-side rudaba gfx + vgl renderer chain and a memory-safe VMM (Rust vmm) to forward GPU ioctls and map host pages into guests. Chromebooks already ship MSM-based support; upstreaming targets Intel, AMD and Apple; NVIDIA and Windows native-context support are not implemented. Benchmarks show about 90% native performance, 98% typical, and native contexts mitigate roughly 80% of past driver vulnerabilities, but denial-of-service, information leaks and firmware bugs remain practical risks.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:49GPU needs for clients
- 3:34GPU importance rising
- 5:00GPU acceleration options
- 8:37Security of API forwarding
- 10:03Verdio GPU native context
- 13:37Security boundaries explained
- 15:25Denial of service risks
- 17:11Information leak risks
- 18:36Privilege escalation risks
- 20:03Security processes
- 20:24Memory mapping in Verdio
- 22:12Libraries and renderer
- 23:59Secure host integration
- 24:43Guest integration options
- 25:49Enterprise use cases
- 26:40Conclusion and Q&A
related talks
01h 00mIsolating GPU access in its own process
35mIsolating GPU access in its own process
33mZephyr‑Based VIRTIO Backend on Xen: Toward Open Source Functional Safety
12mSecurity and Performance in the Delegated User-level Virtualization
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
14mHoneycomb: Secure and Efficient GPU Executions via Static Validation
15mSafeRace: Assessing and Addressing WebGPU Memory Safety in the Presence of Data Races
30mAccelerating AI Securely with GVisor