Memory Read (Standard)

Overview

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

Execution Mechanics

Each thread issues 128-bit (uint4) loads in a coalesced grid-stride loop, unrolled 16 deep, and XORs every value into a register accumulator so the compiler cannot drop the loads.

• The accumulator is written back only when it equals a sentinel it practically never matches, which keeps the sink at zero cost.
• The buffer comes from the runtime allocator, so the vector loads are aligned on both vendors; no decomposition into smaller chunks is needed.

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.