AMD assessed the report as a valid demonstration of resource-contention-based denial-of-service behavior in shared DRAM resources. But Shared DRAM Resource Contention is common, even though in AI/LLM environment.


This article was published on November 19, 2026.

Preface: In normal circumstances, the developers do not write intentional row/bank conflict loops for normal applications—those synthetic loops are strictly for hardware benchmarking and reverse-engineering. However, unintentional row and bank conflicts happen all the time in real-world ROCm workloads when standard software algorithms hit unlucky memory layouts.

Background: The following explains in detail why shared DRAM resource contention issues also occur on GPUs (not limited to CPUs only). Please refer to the accompanying infographic as you read.
In everyday AI, HPC, and graphics workloads running on ROCm, developers never write _mm_clflush or __builtin_nontemporal_load pointer chases. Instead, real operations trigger the exact same hardware stall through normal data structures:
Power-of-Two Matrix Dimensions (GEMM & AI Tensors): Deep learning models frequently use clean power-of-two sizes (such as hidden dimensions of 4096, 8192, or 16384). When a ROCm kernel reads down a column of a matrix instead of across a row, every step jumps by an exact power-of-two byte stride, funneling all parallel threads into the exact same HBM4 bank or LDS partition.
Refer to infographic point 3. The code on left is part of the runnable 64-bit C++/HIP reference program using POSIX fork() to spawn two independent user-space processes sharing the MI455X. Each process allocates its own 64-bit HBM4 buffer and hammers the same HBM4 bank on different rows. Once you’ve gone through points 1 to 7 of the infographic.
After completing points 1 to 7 of the infographic. You may be wondering how ROCm mitigates this situation in production.
• ROCm Libraries Handle It Automatically: Standard ROCm libraries like rocBLAS, hipBLASLt, and Composable Kernel (CK) include pre-tuned tile sizes, memory padding, and XOR address swizzling specifically engineered to prevent HBM and LDS bank conflicts.
• When You Must Manage It Yourself: You only encounter these real-world conflict loops when writing custom HIP or Triton kernels from scratch without padding your shared memory arrays or swizzling your global memory access patterns.

Technical details: AMD-SN-7065 describes the exact same physical theory and hardware mechanism. AMD confirms that an unprivileged user-space workload can cause severe performance degradation (a denial-of-service slowdown) in co-located workloads by intentionally creating targeted DRAM bank contention.

Official announcement: Please refer to the link for details –https://www.amd.com/en/resources/product-security/bulletin/amd-sb-7065.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.