VRAM, short for video random access memory, is the dedicated memory attached to a graphics processor that stores the data the GPU needs at high speed, including textures, geometry, frame buffers and, in compute workloads, model weights and intermediate results. In computer hardware, and within graphics in particular, VRAM capacity and bandwidth are two of the main factors that decide which resolutions, quality settings and AI models a graphics card can handle.
The name dates from early display memory designed for video output, but today the term usually refers to the GDDR or HBM memory on a discrete graphics card. Integrated GPUs normally have no VRAM of their own and borrow a portion of system memory instead.
How VRAM serves the GPU
VRAM exists because a GPU processes enormous volumes of data in parallel and would stall constantly if it had to fetch that data over the comparatively slow link to system memory. Placing fast memory chips directly beside the GPU on the same board, connected by a wide bus, keeps thousands of shader, CUDA and tensor cores supplied with data.
Memory bandwidth is calculated from the bus width and the per-pin data rate: bandwidth in gigabytes per second equals bus width in bits divided by 8, multiplied by the data rate in gigabits per second. A 256-bit bus at 20 Gbps per pin therefore delivers 640 GB/s, which is many times what a typical desktop system memory configuration provides.
The table lists what typically occupies VRAM during graphics and compute work.
| Data type | Workload | What drives its size |
|---|---|---|
| Textures | Games and 3D applications | Texture quality setting and resolution |
| Frame and depth buffers | Rendering | Output resolution and number of buffers |
| Geometry and acceleration structures | Rasterization and ray tracing | Scene complexity |
| Model weights | AI inference and training | Parameter count and numeric precision |
| Activations and key-value cache | AI workloads | Batch size and context length |
Types of VRAM technology
VRAM technologies differ in how they achieve bandwidth: GDDR uses very high per-pin speeds on a moderate bus, while HBM uses a very wide bus at lower speeds with chips stacked vertically next to the GPU. The choice affects cost, power and board design.
| Technology | Approach | Typical use |
|---|---|---|
| GDDR6 | High per-pin rates on 64-bit to 384-bit buses | Mainstream graphics cards |
| GDDR6X | Multi-level signaling for higher rates | Higher-end consumer cards |
| GDDR7 | Newer signaling with further increased rates | Current-generation graphics cards |
| HBM2E, HBM3 and HBM3E | Stacked dies on a 1,024-bit interface per stack | Data center accelerators |
| Shared system memory | Portion of DDR or LPDDR RAM | Integrated GPUs and unified memory chips |
HBM delivers the highest bandwidth, reaching well over a terabyte per second per stack in recent generations, but it requires an interposer and complex packaging. That cost is why HBM appears mainly in AI and high-performance computing accelerators, while GDDR dominates consumer graphics cards.
VRAM compared with system RAM and unified memory
VRAM is optimized for bandwidth, whereas system RAM is optimized for lower latency, larger capacity and lower cost per gigabyte. Data that does not fit in VRAM must move across the PCIe link, which is far slower than local graphics memory.
| Memory type | Main strength | Main limitation |
|---|---|---|
| Dedicated VRAM | Very high bandwidth to the GPU | Fixed capacity, cannot be upgraded |
| System RAM | Large, affordable and upgradable | Much lower bandwidth to a discrete GPU |
| Shared memory on integrated GPUs | No extra memory chips needed | Competes with the CPU for bandwidth |
| Unified memory architectures | CPU and GPU share one large pool | Bandwidth below high-end dedicated cards |
Unified memory designs, used in some ARM-based laptops and workstation chips, allow the GPU to address tens of gigabytes or more without copying. They can run large AI models that exceed the VRAM of consumer graphics cards, though usually at lower token generation speed than a high-end discrete GPU with enough memory.
How much VRAM a workload needs
The right amount of VRAM depends on resolution, texture quality and the size of compute models, not on the GPU’s processing power alone. Requirements also rise over time as new games and models grow.
| Workload | Practical VRAM range | Main driver |
|---|---|---|
| Office work and video playback | Shared memory is sufficient | Display output only |
| 1080p gaming | About 8 GB or more | Texture quality |
| 1440p and 4K gaming | About 12 to 16 GB or more | Resolution, ray tracing, frame generation |
| Professional 3D and video editing | 16 to 48 GB | Scene size, high-resolution timelines |
| Local large language model inference | Depends on parameters and quantization | Model size plus key-value cache |
A useful estimate for AI inference is that model weights need roughly 2 bytes per parameter at 16-bit precision and about 0.5 bytes per parameter with 4-bit quantization, plus extra memory for the context. A 7-billion-parameter model thus needs around 14 GB at 16-bit precision, but fits comfortably in 8 GB when quantized.
Symptoms of insufficient VRAM and common misconceptions
When a workload exceeds available VRAM, the driver moves data to system memory over PCIe, and performance can collapse suddenly rather than decline smoothly. Recognizing the symptoms helps distinguish a memory shortage from a weak GPU.
- Stutter and sharp frame-time spikes at high texture settings often indicate VRAM overflow rather than low GPU speed.
- Textures loading late or appearing blurry can mean the game is reducing quality to stay within memory.
- Out-of-memory errors in AI tools mean the model, batch or context must shrink or be quantized.
- More VRAM does not make a slow GPU fast; it only removes a bottleneck when memory is the limit.
- Reported usage is not true need, because many games allocate spare VRAM as a cache.
In simple terms
VRAM is the graphics card’s own workbench. The graphics processor keeps everything it needs for the current picture, such as textures, shapes and lighting data, laid out on this bench so it can grab items instantly. A bigger and faster bench means less waiting.
If the bench is too small, the processor has to keep fetching items from a storeroom farther away, which is the computer’s main memory. That is when games stutter and AI tools slow down or stop, even if the processor itself is powerful.
A practical example
A small architecture firm wants to run a local AI assistant for drafting project descriptions and also render client walkthroughs. Its current workstation has a graphics card with 8 GB of VRAM. A 13-billion-parameter language model at 16-bit precision needs about 26 GB for weights alone, so it fails with an out-of-memory error, and the 3D walkthroughs stutter whenever high-resolution textures load.
The firm upgrades to a card with 24 GB of VRAM. With 4-bit quantization, the same model now needs about 7 GB for weights and leaves ample room for a long context, while the walkthrough scenes fit entirely in graphics memory. Rendering becomes smooth and the assistant responds quickly, without any change to the processor or system RAM.