NVIDIA 机密计算实现高性能私有 AI 推理
原标题：Enabling Private High-Performance Production AI Inference with NVIDIA Confidential Computing
AI 摘要
NVIDIA 技术博客介绍其机密计算（Confidential Computing）方案，通过内存加密的机密虚拟机、机密 GPU 和加密 NVLink，让 LLM 推理在可信硬件环境中运行。文章以 DeepSeek-R1-0528-NVFP4 模型在 TensorRT LLM 上的测试为例，说明 CC 开启后输出吞吐保留 96.1%–98.2%，平均 TPOT 仅增加 1.2%–4.3%。文中还披露了 TensorRT LLM 针对 CC 环境在内存选择、异步回读、内核自动调优和 NVLS 通信方面的适配措施。
正文节选
As large language model (LLM) inference increasingly processes sensitive information and proprietary model context across personal, enterprise, and regulated settings, data must be processed inside a trusted environment. NVIDIA Confidential Computing (CC) provides a pathway for running these workloads securely using memory-encrypted confidential virtual machines (CVMs), confidential GPUs, and encrypted NVIDIA NVLink. This enables running production AI inference on trusted hardware. Inference fra