HBM Read (Standard)

Overview

The hbm_read test establishes the baseline sequential read bandwidth of the GPU's memory subsystem.

Execution Mechanics

This kernel uses volatile reads decomposed into 32-bit chunks.

• It utilizes a standard 4x unroll per thread.
• Architecture Fix: Attempting to execute single 128-bit vector loads on unaligned buffers can cause segmentation faults on specific AMD RDNA hardware. This kernel decomposes the fetch into smaller, safer chunks to ensure cross-platform stability.

Target Subsystems

• Primary Target: Sequential VRAM Read Bandwidth.

Failure Symptoms

Critical Failures

• Low Throughput: Just like the write test, low throughput signifies memory controller degradation or ECC intervention.