Interposer-Based Root of Trust

Tapojyoti Mandal, Gino Chacon, Johann Knechtel, Ozgur Sinanoglu, Paul Gratz, Vassos Soteriou

Abstract

Industry is moving towards large-scale systems where processor cores, memories, accelerators, etc. are bundled via 2.5D integration. These various components are fabricated separately as chiplets and then integrated using an interconnect carrier, a so-called interposer. This new design style provides benefits in terms of yield as well as economies of scale, as chiplets may come from various third-party vendors, and be integrated into one sophisticated system. The benefits of this approach, however, come at the cost of new challenges for the system's security and integrity when many third-party component chiplets, some from not fully trusted vendors, are integrated. Here, we explore these challenges, but also promises, for modern interposer-based systems of cache-coherent, multi-core chiplets. First, we introduce a new, coherence-based attack, GETXspy, wherein a single compromised chiplet can expose a high-bandwidth side/covert-channel in an ostensibly secure system. We further show that prior art is insufficient to stop this new attack. Second, we propose using an active interposer as generic, secure-by-construction platform that forms a physical root of trust for modern 2.5D systems. Our scheme has limited overhead, restricted to the active interposer, allowing the chiplets and the coherence system to remain untouched. We show that our scheme prevents a wide range of attacks, including but not limited to our GETXspy attack, with little overhead on system performance, ~4%. This overhead reduces as workloads increase, ensuring scalability of the scheme.

Introduction

As Moore’s Law slows down, the semiconductor industry is pivoting toward 2.5D integration. Instead of building one massive, expensive "monolithic" chip, designers are now bundling smaller, specialized "chiplets" (cores, memory, accelerators) onto a shared communication highway called an interposer. While this "Lego-like" approach boosts yields and lowers costs, it introduces a massive security headache: How do you trust a system when the individual pieces come from different, potentially untrusted third-party vendors?

In a multi-chiplet system, all components must maintain a consistent view of memory—a process called cache coherence. If Chiplet A wants to write to a memory address that Chiplet B has cached, the system sends a "coherence message" to invalidate Chiplet B’s copy. Current security defenses focus on the Network-on-Chip (NoC) or software layers, but they largely ignore the "legal" but exploitable behavior of the coherence protocol itself. Because these protocols are predictable and broadcast messages across the system, a malicious chiplet can "snoop" on traffic it shouldn't see or even forge messages to trick the system into handing over sensitive data.

The GETXspy Attack

The researchers identified a critical gap in current defenses and developed GETXspy, a novel coherence-based covert-channel attack. In this scenario, a "spy" process on a clean chiplet performs specific memory writes that trigger "Get Exclusive" (GETX) coherence requests. A "Trojan" hidden in a different, compromised chiplet observes these broadcasted invalidation messages. By timing or sequencing these legal requests, the spy can leak data to the Trojan at a rate of 4.22 Mbps without ever violating standard system rules.

Overview of the GETXspy coherence-based attack mechanism.

This attack is particularly potent because it stays perfectly "legal" within the protocol's rules, making it invisible to traditional NoC-level or software-based security monitors.

Active Interposer Root of Trust

To counter such threats, the authors propose moving security logic off the chiplets and onto the active interposer. Since the interposer is the physical backbone connecting all chiplets, every single message must pass through it. This makes it an ideal location for a hardware-enforced "Root of Trust."

The proposed architecture implements:

  • Coherence Message Checkers (CMCs): Specialized hardware filters placed at the entry points (ingress) of the interposer.
  • Address Protection Unit (APU) Table: A secure lookup table that stores memory permissions (Read-Only, Read-Write, No Access) for every chiplet.
  • CMC-1 & CMC-2: CMC-1 validates that incoming requests from chiplets match their assigned permissions. CMC-2 intercepts broadcast messages from memory controllers and "diverts" them away from chiplets that don't have permission to see that memory region, effectively blinding the Trojan.
Architecture of the Active Interposer Layer featuring Coherence Message Checkers (CMCs).

By placing the defense in the interposer, the system doesn't need to trust the chiplets or the vendors who made them. The interposer acts as a physical gatekeeper, ensuring zero-modification coherence—meaning the underlying protocols (e.g., MOESI) remain untouched, avoiding complex deadlock issues.

Evaluation and Results

The design was evaluated using the gem5 system simulator, modeling an 8-chiplet, 64-core system similar to modern high-performance processors. Performance was measured using the SPEC 2006 benchmark suite.

Key results include:

  • Security Coverage: The proposed CMC architecture successfully blocked GETXspy, as well as other common threats like masquerading (spoofing a different chiplet's ID) and diverting (rerouting data to an unauthorized chiplet).
  • Low Overhead: The average performance hit was only ~4%. Interestingly, as the system workload increases, the relative overhead decreases, suggesting the design scales well for massive data-center applications.
  • Attack Potency: Without defense, GETXspy achieved a bandwidth of 4.22 Mbps, sufficient to leak encryption keys in milliseconds.

Conclusion

As the industry moves toward a "chiplet-first" world, this research provides a vital blueprint for building secure systems out of potentially insecure parts. By leveraging the active interposer as a physical root of trust, designers can enforce security at the micro-architectural level, protecting against sophisticated coherence-based attacks with minimal performance impact.

Summary updated

Abstract

Industry is moving towards large-scale systems where processor cores, memories, accelerators, etc.\ are bundled via 2.5D integration. These various components are fabricated separately as chiplets and then integrated using an interconnect carrier, a so-called interposer. This new design style provides benefits in terms of yield as well as economies of scale, as chiplets may come from various third-party vendors, and be integrated into one sophisticated system. The benefits of this approach, however, come at the cost of new challenges for the system's security and integrity when many third-party component chiplets, some from not fully trusted vendors, are integrated. Here, we explore these challenges, but also promises, for modern interposer-based systems of cache-coherent, multi-core chiplets. First, we introduce a new, coherence-based attack, GETXspy, wherein a single compromised chiplet can expose a high-bandwidth side/covert-channel in an ostensibly secure system. We further show that prior art is insufficient to stop this new attack. Second, we propose using an active interposer as generic, secure-by-construction platform that forms a physical root of trust for modern 2.5D systems. Our scheme has limited overhead, restricted to the active interposer, allowing the chiplets and the coherence system to remain untouched. We show that our scheme prevents a wide range of attacks, including but not limited to our GETXspy attack, with little overhead on system performance, $\sim$4\%. This overhead reduces as workloads increase, ensuring scalability of the scheme.
BibTeX
@misc{mandal2021interposer,
  author = {Tapojyoti Mandal and Gino Chacon and Johann Knechtel and Ozgur Sinanoglu and Paul Gratz and Vassos Soteriou},
  title = {{Interposer-Based Root of Trust}},
  year = {2021},
  eprint = {2105.02917},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.02917},
}
Figure