Do not contempt ZENeath (Exploiting Contention-Channel Vulnerabilities on Chiplet-Based AMD Zen Silicon). Its severity depends on how the vulnerability is exploited in a specific environment (10th Nov 2026)


Preface: The ZENeath paper demonstrates that contention can be monitored on the internal interconnect structures (the Infinity Fabric inside the Zen chip packages).

Background: Attached infographic was originally designed to focus on scheduler contention-side channel attacks on AMD “Zen” processors, but it also applies to “ZENeath – Exploiting the Contention-Side Channel Vulnerability”. The following is a detailed explanation. The following explanation requires the use of the accompanying infographic. In orchestration brain topology, when a Zen CPU communicates with a GPU (whether an H100 or an MI455X), that data cannot remain localized inside a single CPU core. It must be pushed out via the host system buses. Because the CPU’s internal fabric is the gateway to those external links, the internal contention channel acts like a window into external communication. Even though the data ultimately leaves the AMD ecosystem to go to an NVIDIA GPU via PCIe Gen 6, it must first travel through the Zen chip’s internal I/O Die (IOD) to reach the root complex controllers.

•If your orchestration logic executes The Vulnerable Way (secret-dependent data copy sizes), the IOD will process variable amounts of traffic traveling toward Route Option A.

•ZENeath proves that an attacker sharing that internal AMD infrastructure can spy on these variations. Therefore, even though Route Option A goes to an external device, the secret-dependent characteristics of that external transfer are leaked internally before the data ever leaves the CPU.

How it Extends to Route Option B (AMD Instinct Cores)

For Route Option B, AMD blends the boundaries between internal and external communication.

•AMD Instinct GPUs natively connect via external Infinity Fabric / xGMI links.

•This means the exact same routing mechanisms, bus protocols, and hashing logic that ZENeath exposes inside the chip are extended outside the chip across the system motherboard to the GPU.

•If your code handles encrypted model weights using variable copy lengths down to Route Option B, you are essentially broadcasting that timing signature across the entire unified fabric.

Conclusion: Attached infographic successfully highlights the danger of core-level scheduler contention. Bringing ZENeath into the equation does not invalidate this diagram—it scales it up. It proves that if a developer ignores advice in Point 3 (Fixed Size/Constant-Time), the resulting timing signatures do not just leak to a neighboring thread; they propagate through the entire fabric infrastructure, compromising secure data orchestration across the host CPUs and the AI accelerators.

Official announcement: Please refer to the link for details –

https://www.amd.com/en/resources/product-security/bulletin/amd-sb-7051.html


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.