> ## Documentation Index > Fetch the complete documentation index at: https://docs.phala.com/llms.txt > Use this file to discover all available pages before exploring further. > Performance benchmarks and metrics for LLM inference in GPU TEE environments. # Benchmark ## Key Results **TEE mode on H100/H200 GPUs runs up to 99% efficiency, nearly matching native performance.** This means you get confidential computing with minimal performance penalty. The efficiency improves as your workload grows. Larger models and longer sequences see even better performance relative to native execution. ## What We Tested We benchmarked LLM inference in GPU TEE mode using NVIDIA H100 and H200 GPUs. TEE mode encrypts your data and computation while it runs on the GPU. Performance comparison table showing TPS and QPS metrics for LLama-3.1-8B, Phi3-14B-128k, and LLama-3.1-70B models on H100 and H200 GPUs with TEE-on vs TEE-off overhead percentages ## Performance Insights * **Larger models perform better in TEE mode.** Models like Phi3-14B-128k and Llama3.1-70B show higher efficiency because they spend more time computing and less time on I/O operations. * **Longer sequences boost efficiency.** When you process more tokens (input + output), the computation-to-I/O ratio improves. This makes TEE overhead less noticeable. * **The sweet spot: High-computation workloads.** TEE mode shines when your GPU is doing heavy lifting. The encryption/decryption overhead becomes trivial compared to the actual computation time. ## Resources For detailed metrics and analysis, check our [benchmark paper](https://arxiv.org/pdf/2409.03992). See detailed [GPU TEE performance specifications](https://phala.com/gpu-tee) for production deployments, including [H100 TEE](https://phala.com/gpu-tee/h100) and [H200 TEE](https://phala.com/gpu-tee/h200) configurations.