Confidential AI Computing Hits Production
Corvex has achieved a verified production deployment of confidential computing for AI on NVIDIA's new HGX B200 systems. The technology enables secure, encrypted processing of sensitive video and data at GPU scale. This is a critical development for newsrooms handling legal, medical, or investigative footage who need to meet strict data privacy requirements.
Confidential computing creates a hardware-based trusted execution environment (TEE), or secure enclave, that isolates data and the code processing it from the rest of the system. This means data remains encrypted even while in use, protecting it from the operating system, hypervisor, and even administrators of the cloud or edge server where the computation is happening. The market for this technology is projected to grow from $5.3 billion in 2023 to $59.4 billion by 2028. Corvex's deployment pairs Intel's Trust Domain Extensions (TDX) for the CPU with NVIDIA's Confidential Computing capabilities for the GPU, providing end-to-end protection. This addresses the entire compute stack, isolating workloads and ensuring data remains encrypted as it moves between the CPU and the powerful NVIDIA B200 GPUs. The system uses cryptographic proof, called remote attestation, to verify the integrity of the hardware and software before processing begins. The NVIDIA HGX B200 platform is built on NVIDIA's Blackwell architecture, announced in March 2024, and features eight B200 Tensor Core GPUs. Each B200 GPU provides up to 18 PFLOPS of FP4 tensor processing power and is equipped with 192GB of HBM3e memory, delivering 8 TB/s of memory bandwidth for handling massive datasets and complex AI models. This architecture offers up to 15 times faster inference performance than the previous Hopper generation. This level of security has historically come with a performance trade-off, but NVIDIA's Blackwell architecture is designed to deliver near-native performance for encrypted workloads. The HGX B200 platform also focuses on energy efficiency, offering up to 12 times lower energy consumption compared to the prior generation, a key consideration for scaling video processing infrastructure. For newsrooms, this technology directly addresses the risks associated with analyzing AI-driven video content, which can contain biometric data or other sensitive information subject to regulations like GDPR. It provides a verifiable way to protect proprietary model weights and sensitive source material from both internal and external threats, which is crucial when collaborating with third parties or using multi-tenant cloud environments.
