Answers are grounded in this paper and cite their sources.
Cloud providers increasingly deploy accelerators like GPUs and TPUs for AI and data analytics. To secure these devices, emerging trusted I/O architectures (e.g., Intel TDX-Connect, AMD SEV-IO) extend the host CPU’s trusted domain to the accelerator fabric. However, these CPU-centric designs have three limitations: (1) TCB bloat: any guarantee about where data flows rests on a full guest OS; (2) Coarse-grained control: lacking application-aware data flow policies; and (3) Centralized bottlenecks: host-mediated peer-to-peer traffic puts the host on every transfer’s data path. To address these challenges, we propose a decentralized trusted bus that provides attested, hardware-protected channels for direct data movement between accelerators. IronBus realizes this abstraction through a cross-layer hardware/OS/compiler co-design for next-generation interconnect fabrics. IronBus shifts access control from the centralized host to a per-device Attested Interface Unit (AIU) —a minimal silicon Root-of-Trust that enforces security policies at the fabric endpoints. Orchestrating these units is a Hardware-Abstraction Layer (HAL) —a hardware-proximate lean software layer with formally verified protocols that replaces the guest OS to bootstrap trust. IronBus also introduces a DFG-based programming model and policy compiler : developers express multi-accelerator pipelines as Device Flow Graphs (DFGs) , from which the compiler automatically synthesizes Security Service Level Agreements (SSLAs) , and the HAL signs execution proofs for auditing. We prototype IronBus on a many-core RISC-V platform, evaluate it in gem5 , and formally verify its protocols with Tamarin and CryptoVerif . The IronBus HAL is as small as the SEV and TDX firmware, whereas for their data-flow policy host-centric designs trust a guest OS and firmware that are together 3–94 × larger. Direct device-to-device channels make IronBus 1.1–2 × faster than host-centric designs on accelerator chains, collectives, and ML workloads, even when every device has its own host link. The cryptographic bus costs 5–23% on communication microbenchmarks and at most 10% on ML workloads over an unprotected baseline.
No takes yet. Share an insight, caveat, or question.
Unnibhavi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: