Graphics Card Specifications Explained: VRAM, Bandwidth and Power
Some links in my articles are affiliate links. I may earn a commission when you buy through them.
Use the specification sheet to check capability and compatibility, then use relevant performance tests to compare speed. Memory capacity, memory bandwidth and processor clock describe different things; none can rank every graphics card on its own.
Record the full GPU name, memory capacity and partner model. A product family can include different memory capacities or configurations. A cooler name such as Dual or Gaming also belongs to a particular card maker rather than describing a universal performance level.
Open the GPU maker’s specification and the card maker’s page. The first gives the architecture and core features; the second gives dimensions, outputs, power connections and the supplied cooler. If a shop title is shortened, use the exact product code to identify what is being sold.
For the overall choice, start with the graphics-card guide. The table below is the checklist I would use once two or three cards are already worth comparing.
| Specification | Useful meaning | Purchase check |
|---|---|---|
| VRAM capacity | How much local graphics data can be held. | Enough for games, textures or projects. |
| Memory bandwidth | Rate at which the memory subsystem can transfer data. | Compare with architecture and workload results. |
| Core/boost clock | Operating frequency of that design. | Not a cross-brand speed ranking. |
| Compute units or cores | Resources within a particular architecture. | Different vendors count different building blocks. |
| Board power | The card’s power provision and cooling class. | Exact PSU and connector requirements. |
| Outputs and dimensions | Physical integration with the rest of the setup. | Cable capabilities, case fit and clearance. |
VRAM size and bandwidth are separate
Capacity is measured in GB. Bandwidth is commonly quoted in GB/s and depends on the memory data rate and bus width. A wider bus or newer memory type can increase theoretical bandwidth, but caches and the GPU architecture affect how efficiently that bandwidth is used.
A simple bandwidth calculation multiplies effective memory data rate by bus width, then divides by eight to convert bits to bytes. For example, a hypothetical 20Gb/s memory interface across 128 bits gives 320GB/s. That is arithmetic for the interface, not a benchmark result or a claim that one real card is faster.
Use the VRAM capacity guide when deciding how much memory you need. Compare the same game or application on the exact cards when judging the memory system’s effect.
Why cores, GHz and TFLOPS do not settle the ranking
A GPU’s core count describes execution resources under that vendor’s architecture. A CUDA core, an AMD stream processor and an Intel Xe-core are not interchangeable units. Clock speed also needs context: a higher frequency on a smaller or older design does not establish faster rendering.
TFLOPS describes a theoretical floating-point throughput under stated assumptions. A game also depends on texture processing, memory access, geometry, scheduling and specialised hardware. Professional workloads use different instructions and data types, so a headline compute number may describe a capability your application never uses.
When a comparison lists GPU TOPS, check the numerical precision and whether sparsity is included. That number is useful only alongside the supported AI workload and its memory requirement. The local-AI shortlist keeps those practical constraints visible.
Power numbers belong with the whole PC
Board-power figures and recommended system-PSU ratings describe different things. The PSU recommendation includes an assumed complete system; it is not the GPU’s measured consumption. Partner cards can also specify different power limits or connectors from a reference design.
Record the exact card’s power leads and the PSU’s supported cable arrangement. Modular PSU-side plugs are not universal, even where a cable’s GPU-side plug looks familiar. Follow the PSU and GPU documentation for approved cables, seating and routing. The upgrade checklist puts those checks alongside installed space.
For a quiet build, compare cooler behaviour at a useful load and noise level. A slightly faster factory clock can be a poor reason to buy a physically awkward or much more expensive model. The compact/quiet shortlist starts with that trade-off.
Turn the specification into a buying comparison
Keep a short record of exact model, capacity, useful software features, installed dimensions and total delivered upgrade cost. Compare clearly labelled benchmarks for the workload that motivated the purchase. A specification rules unsuitable products out; the workload comparison chooses between the suitable ones.
Use the GPU buying guide to find the relevant use-case shortlist and the PCIe compatibility guide if the card is going into an older system.
Intel graphics memory, bandwidth and compute specifications · NVIDIA’s memory-subsystem explanation