Benchmarking Confidential Computing Performance on Nvidia Blackwell GPUs (arxiv.org)

A recent study has benchmarked the performance of confidential computing on NVIDIA's Blackwell GPUs, specifically focusing on large language model inference and training within a Trusted Execution Environment (TEE). Utilizing Intel's Trust Domain Extensions (TDX) with NVIDIA's Confidential Computing technology, the researchers compared confidential and non-confidential runs to assess performance impacts. The findings revealed that when properly configured, confidential inference on Blackwell GPUs can achieve a throughput overhead of just 1-3%, significantly lower than the 30-40% penalties observed with suboptimal configurations. This research is significant for the AI/ML community as it demonstrates that secure computing environments can be employed with minimal performance trade-offs, which is crucial for applications requiring data privacy, such as healthcare and finance. The study also identifies key cost factors associated with encrypted operations, providing insights into how batch sizes and data traffic can influence overall performance. Importantly, it reassures stakeholders that the core capabilities of GPU computing, including energy efficiency and memory capacity, remain unaffected by the adoption of confidential computing technologies.