Honeycomb: Secure and Efficient GPU Executions via Static Validation
summary
Honeycomb is a software GPU TEE that uses static validation of GPU binaries plus OS-level security monitors to remove the OS and driver from the TCB, prototype targets an AMD RX6900XT, and co-designs with a CPU TEE (AMD SEV-SNP). The system validates 23 applications across five benchmarks, uses polyhedral analysis for kernel memory-range checks, implements fast plaintext shared-device-memory IPC, and shows an 18× smaller TCB, 529× faster IPC, and about 2% overhead on BERT and NanoGPT, demonstrating a practical, working prototype.
AI-written summary · about
chapters (10)
- 0:00Honeycomb intro and goals
- 1:56GPU TEE challenges
- 3:01Honeycomb’s GPU TEE design
- 5:28GPU assumptions and basics
- 6:32Honeycomb architecture
- 8:17Security monitors implementation
- 8:57Static validation principles
- 11:09Efficient IPC validation
- 11:53Evaluation results
- 13:15Conclusion and future
related talks
41mPanel: A New Root Of Trust: Changing Computing With TEE
23mSAGE: Software-based Attestation for GPU Execution
42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
36mEverybody is talking about Confidential Computing - the minimum Devs should know- 42mSponsored Session: Confidential Computing - New Capabilities for New Workloads
30mProtecting Wasm Workloads with Confidential Computing and TEE...
07mBuilding Robust GPU Performance Validation into CI/CD Pipelines
44mGPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud