All guides

Memory specifications

GPU memory explained: VRAM, bus width and bandwidth

Learn what GPU VRAM capacity, memory type, bus width and bandwidth describe and why none is a complete performance score.

Sources and review

Direct answer

VRAM capacity describes local memory space. Bus width describes the width of the memory interface and bandwidth describes a transfer rate. Cache behavior and the application's memory access pattern affect the rate actually achieved. [1][2]

Check it in order

  1. 1Check the application's documented memory requirements and limits. [1]
  2. 2Confirm the capacity of the exact GPU variant because one model name can cover multiple versions. [3]
  3. 3Read memory type, bus width and transfer rate together rather than ranking one value alone. [2][3]
  4. 4Use workload tests with the scene or settings that matter to you. [1]
  5. 5Keep a memory-capacity limit separate from a measured throughput result. [1][2]

VRAM capacity

Capacity matters when a scene must fit in GPU memory. Blender 4.5 documents attempts to use system memory when local memory fills on supported backends, with a performance cost. It also documents out-of-memory errors and smaller textures as a way to reduce scene memory use. [1]

Do not turn capacity into a speed rating. Whether a scene fits in memory and how fast it renders are separate questions; Blender notes that rendering speed depends on the hardware, technology and scene. [1]

Bus width and memory type

Bus width is the memory interface width expressed in bits. NVIDIA's theoretical-bandwidth calculation combines interface width with memory clock rate and the memory's data-rate factor. A wider bus alone is not an application result. [2]

Cache reuse and access patterns also matter. NVIDIA's CUDA guidance explains that on-chip L2 cache can offer higher bandwidth and lower latency than external memory access. [2]

Memory bandwidth

Theoretical bandwidth describes a hardware transfer limit. Measured bandwidth depends on the program. Neither is a gaming frame-rate or rendering score. [2]

Preserve the manufacturer's label. AMD lists Memory Bandwidth and Effective Memory Bandwidth separately for the RX 7900 XTX. Do not silently compare these as if they were the same metric. [4]

Variant scope matters

NVIDIA's RTX 3080 specifications list 12 GB and 10 GB configurations with different interface widths. Check the capacity and full board model rather than assuming the first value applies to every card under that name. [3]

What the numbers cannot prove

  • VRAM capacity is not a speed rating. [1]
  • Bus width alone does not describe the complete memory subsystem. [2]
  • Theoretical bandwidth does not guarantee application throughput. [2]

Sources and review

SpecMatchup editorial method

This guide is a synthesis of the cited manufacturer documentation and workload methodology, not firsthand hardware testing. SpecMatchup uses these sources to explain practical checks. Product-specific examples do not establish compatibility for every system.

Substantive claim review: . A successful link check alone does not change this date.

Reviewed capacity limits, interface-width terminology and theoretical versus measured bandwidth. Scoped memory-overflow behavior to Blender 4.5 and kept AMD's effective-bandwidth label distinct.

  1. Blender 4.5 LTS Manual: GPU rendering

    Blender - Primary source

    Rendering technologies, memory-limit FAQs and out-of-memory guidance. Backend performance varies with hardware and the scene.

  2. CUDA C++ Best Practices Guide: bandwidth and memory

    NVIDIA - Primary source

    Sections 9.2, 10.1 and 10.2.2 distinguish theoretical from measured bandwidth, host transfers and L2 cache behavior. CUDA examples are not gaming benchmarks.

  3. GeForce RTX 3080 family specifications

    NVIDIA - Primary source

    Memory specifications list separate RTX 3080 capacities and interface widths. Shipping partner specifications must be checked separately.

  4. Radeon RX 7900 XTX specifications

    AMD - Primary source

    Requirements and GPU Memory sections. Separates Typical Board Power from the minimum PSU recommendation and raw bandwidth from effective bandwidth.