Abstract: Modern data centers have grown beyond CPU nodes to provide domain-specific accelerators such as GPUs and FPGAs. From a security standpoint, cloud customers want to protect their data using trusted execution environments (TEEs), but currently must choose between performance and security. This paper proposes a hybrid design that enables data-center-scale confidential computing across heterogeneous nodes, including those without native TEE support. By utilizing a centralized Security Controller and custom hardware primitives, the solution achieves marginal performance overheads of 0.42% to 8% on real-world AI workloads, demonstrating a scalable path for future trusted computing.
Introduction
Modern data centers are no longer just racks of CPUs; they are heterogeneous environments packed with GPUs, FPGAs, AI accelerators, and smart storage. While robust security solutions like Intel SGX or AMD SEV exist for CPUs, these "enclaves" do not easily extend to the specialized accelerators that drive modern workloads.
Currently, cloud tenants face a frustrating choice: prioritize performance by using high-speed accelerators while leaving data exposed, or prioritize security by using CPU-based confidential computing and sacrificing specialized hardware speedups. Making every device TEE-capable is a long-term goal, but connecting them securely across a complex data center fabric without creating bottlenecks is an immediate architectural hurdle.
The Hybrid TEE Architecture
The authors propose a Hybrid TEE design that scales in two dimensions: it handles legacy "non-TEE" nodes and accommodates future TEE-enabled devices. This is achieved through several innovative components:
- The Security Controller (SC): For legacy nodes, a trusted hardware module called the SC acts as a "TEE proxy." It sits on the communication path to enforce isolation, manage attestation, and provide secure channels.
- Hardware Primitives for Accelerators: For new AI accelerators and SSDs, two key components were added: the Access Control Unit (ACU) for memory isolation and the Memory Protection Engine (MPE) for transparent encryption and integrity protection.
- Hierarchical Property-Based Attestation (PBA): A novel tree-like structure used to verify that a device is in a safe state, checking complex configurations rather than just a single hash.
Evaluation and Key Results
The system was evaluated using AI workloads (CIFAR-10, ImageNet) and storage benchmarks (FIO). Hardware costs were estimated using cycle-accurate simulators and RTL synthesis (28nm CMOS).
- Low Performance Overhead: Real-world AI workloads showed a marginal performance hit of only 0.42% to 8%.
- Scalability: A single multi-core SC can support up to 912 concurrent nodes running inference at full bandwidth.
- Minimal Footprint: The hardware additions for the AI accelerator consumed only 0.2% of the die area and 0.38 µW of power.
- Data Pipelining: While encryption adds latency to DMA transfers, the authors used pipelining to hide most of this cost, ensuring efficient data movement.
Insights and Future Directions
A key insight of this research is the "TEE Proxy" concept, which allows for a trusted perimeter around legacy devices. Additionally, the authors observed that the static memory requirements of AI kernels can be leveraged to simplify security architectures.
While the current design assumes physical tamper-resilience for the SC and relies on static memory allocation, future work will explore dynamic allocation and address side-channel vulnerabilities. This hybrid approach provides a practical, scalable path toward comprehensive confidential computing in the modern data center.
Summary updated