AI Practices AI实践 8d ago Updated 7d ago 更新于 7天前 49

Hardware-Rooted AI Security That Won’t Slow You Down 不会拖慢速度的硬件级AI安全

NVIDIA Confidential Computing (CC) on Blackwell GPUs achieves up to 98% performance retention compared to non-secure modes, addressing the historical trade-off between security and speed. The solution utilizes hardware-rooted trust via fused private signing keys, NVLink encryption, and remote attestation through the NVIDIA Remote Attestation Service (NRAS) to protect data and models during inference. Significant performance optimizations include CC-safe autotuning in FlashInfer, async device-to- NVIDIA Confidential Computing (CC) 在 Blackwell GPU 上实现了硬件级安全,保护推理过程中的数据、代码和模型完整性。 启用 CC 后的吞吐量与延迟开销极小,性能保持在无安全方案时的 98% 以上(通常低于 8% 的损耗)。 通过 FlashInfer 的 CC 安全自动调优、SGLang 的异步拷贝及分段 CUDA 图支持等技术优化,缓解了加密带宽和提交延迟的影响。 远程证明(NRAS)仅在启动时进行一次,运行期间不增加单次推理请求的延迟,确保生产环境的高可用性。 该方案解决了 AI 采用中的数据隐私和主权顾虑,使企业能在合规前提下大规模部署机密

65
Hot 热度
70
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • NVIDIA Confidential Computing (CC) on Blackwell GPUs achieves up to 98% performance retention compared to non-secure modes, addressing the historical trade-off between security and speed.
  • The solution utilizes hardware-rooted trust via fused private signing keys, NVLink encryption, and remote attestation through the NVIDIA Remote Attestation Service (NRAS) to protect data and models during inference.
  • Significant performance optimizations include CC-safe autotuning in FlashInfer, async device-to-host copy workers in SGLang, and piecewise CUDA graph support to mitigate encryption overheads.
  • Benchmarks on the HGX B300 with the Qwen 3.5-397B-A17B-FP8 model show minimal throughput and latency overhead (typically under 8%) across various concurrency levels and batch sizes.

Why It Matters

This development removes a major barrier to enterprise AI adoption by demonstrating that rigorous hardware-level security does not require sacrificing inference performance. For organizations handling sensitive data or proprietary models, this provides a viable path to regulatory compliance and data sovereignty without impacting operational efficiency. It signals a maturation of confidential computing technologies, making them practical for production-scale, high-throughput AI deployments.

Technical Details

  • Hardware Root of Trust: Blackwell GPUs (RTX PRO 6000, HGX B200, HGX B300) embed CC in silicon with fused private signing keys that are never exposed to software or firmware, establishing an immutable attestation chain.
  • Remote Attestation: The NVIDIA Remote Attestation Service (NRAS) verifies GPU hardware reports and CPU TEE measurements (AMD SEV-SNP or Intel TDX) against a Reference Integrity Manifest (RIM) before deploying secrets like model decryption keys.
  • Performance Mitigations: Key innovations include replacing event timers with GPU global timer registers in FlashInfer for accurate autotuning, moving token readback off the critical path in SGLang to restore compute/copy overlap, and implementing piecewise CUDA graph replay to reduce kernel launch overhead.
  • Benchmark Results: Testing on Qwen 3.5-397B-A17B-FP8 showed relative performance deltas ranging from -1.0% to -8.1% depending on concurrency and input/output sequence lengths, with most scenarios staying within the 2-5% overhead range.

Industry Insight

  • Adoption Acceleration: Enterprises previously hesitant to deploy AI due to data privacy concerns can now adopt confidential computing with negligible performance penalties, likely accelerating deployment in regulated industries like finance and healthcare.
  • Framework Integration: The close collaboration between NVIDIA and open-source inference frameworks (FlashInfer, SGLang) suggests that future AI infrastructure will increasingly prioritize security-by-design optimizations rather than treating security as an afterthought.
  • Standardization of Attestation: The integration of NRAS with standard CPU TEEs establishes a unified verification protocol, potentially becoming an industry standard for verifying the integrity of AI workloads in multi-vendor environments.

TL;DR

  • NVIDIA Confidential Computing (CC) 在 Blackwell GPU 上实现了硬件级安全,保护推理过程中的数据、代码和模型完整性。
  • 启用 CC 后的吞吐量与延迟开销极小,性能保持在无安全方案时的 98% 以上(通常低于 8% 的损耗)。
  • 通过 FlashInfer 的 CC 安全自动调优、SGLang 的异步拷贝及分段 CUDA 图支持等技术优化,缓解了加密带宽和提交延迟的影响。
  • 远程证明(NRAS)仅在启动时进行一次,运行期间不增加单次推理请求的延迟,确保生产环境的高可用性。
  • 该方案解决了 AI 采用中的数据隐私和主权顾虑,使企业能在合规前提下大规模部署机密 AI 推理。

为什么值得看

对于 AI 从业者而言,这篇资讯打破了“安全必然牺牲性能”的传统认知,证明了硬件级机密计算可以在几乎不影响推理速度的情况下提供企业级安全保障。它为解决数据隐私、主权合规以及模型权重保护提供了切实可行的工程化路径,是构建可信 AI 基础设施的关键参考。

技术解析

  • 硬件信任根与远程证明:Blackwell GPU(如 HGX B200/B300)在制造时融合私有签名密钥,结合 CPU TEE(AMD SEV-SNP 或 Intel TDX)测量,通过 NVIDIA 远程证明服务 (NRAS) 验证系统完整性。一旦验证通过,密钥被注入机密虚拟机 (CVM),后续推理无额外延迟。
  • NVLink 加密与多卡支持:CC 支持多达 8 张 GPU 之间的 NVLink 加密通信,确保跨卡数据传输的安全性,同时保持硅片、互连和系统软件层面的端到端保护。
  • 性能优化创新:针对 CC 模式下的挑战,FlashInfer 使用 GPU 全局定时器替代事件计时器以实现准确的自动调优;SGLang 引入异步 D2H 拷贝工作线程和分段 CUDA 图回放,减少内核启动开销并恢复计算与拷贝的重叠,从而抵消加密带宽限制带来的影响。
  • 基准测试结果:在 HGX B300 上使用 Qwen 3.5-397B-A17B-FP8 模型进行测试,不同并发数(4-256)和输入/输出序列长度下,吞吐量下降幅度介于 -1.0% 至 -8.1% 之间,中位 TPOT 延迟增加也在可接受范围内,证实了接近原生的性能表现。

行业启示

  • 合规即竞争力:随着全球对 AI 数据隐私监管趋严,具备原生硬件级机密计算能力的平台将成为金融、医疗等高敏感行业首选,企业应优先评估基础设施的安全合规能力。
  • 安全与性能的平衡点已突破:传统认为的安全开销已成为历史,新一代 GPU 架构和软件栈优化使得“零信任”架构可以无缝集成到高性能 AI 推理流水线中,无需为安全妥协业务 SLA。
  • 生态协同至关重要:NVIDIA 与 FlashInfer、SGLang 等主流推理框架的深度合作表明,机密计算的性能优势依赖于软硬件栈的整体优化,开发者需关注框架对 CC 模式的适配情况以最大化部署效益。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Security 安全 GPU GPU Inference 推理