← All resources

Learning About GPUs Through Measuring Memory Bandwidth

Auto-discovered

Blog Advanced

Manon Oomen

Visit resource ↗

Description

Learning About GPUs Through Measuring Memory Bandwidth is a Traverse Research / Evolve blog article by Manon Oomen on designing microbenchmarks that isolate GPU load and store bandwidth across caches and VRAM.

It reviews descriptors, byte-address vs structured vs typed buffers, texture-unit work, cache hierarchy and write policies, then walks shader iterations that amortize dispatch overhead, avoid false cache hits, and compare buffers versus textures and element widths. Findings cover Adreno 740 on Quest 3, Radeon RX 9070 XT float vs int ALU traps near a fast L0, Intel Arc B580 element-size quirks, and GeForce RTX 5070 Ti write-contention oddities.

It is for advanced graphics programmers who already write compute shaders and want hardware-aware bandwidth intuition. HLSL-style examples assume DirectX familiarity.

The post is a free long-form essay with diagrams, shader listings, and charts.

It is useful when interpreting bandwidth cliffs in profiling and deciding whether buffer, texture, or element-size choices matter for a target GPU.