The short version
- VRAM (video RAM) is memory that sits on the graphics card itself. It holds everything the GPU needs to draw a frame: textures, models, lighting data and the frames being built.
- Two numbers matter: capacity (how much fits, in GB) and bandwidth (how fast the GPU can read it, in GB/s).
- When a game needs more VRAM than the card has, data spills into much slower system RAM. You get stutter, missing or blurry textures, and sometimes crashes.
- In 2026, 8GB is tight even at 1080p in demanding games. 12GB is the floor for 1440p, and 16GB is the comfortable choice.
What VRAM actually stores
Your PC has two kinds of memory for games. System RAM (DDR4 or DDR5) feeds the CPU, which runs the game's logic, AI and physics. VRAM feeds the GPU, which turns all of that into pixels, and it's soldered right next to the graphics chip so data has the shortest possible trip.
Here's what fills it up:
- Textures, the images wrapped around every surface. These are usually the biggest item, and the "Texture quality" setting mostly decides how much VRAM a game uses.
- Geometry, the 3D models of characters, buildings and terrain.
- Frame buffers, the frames being drawn plus extra buffers for depth, lighting and post-processing. Higher resolutions need bigger buffers.
- Ray tracing data, a structure called a BVH that maps every object in the scene so the GPU can work out where light rays hit. Turning ray tracing on can add a gigabyte or more.
- Upscaling and frame generation. DLSS, FSR and frame generation keep previous frames and motion data in memory to build new ones. Our DLSS explainer covers how that works.
A game streams data in and out of VRAM constantly as you move through the world. As long as everything the current scene needs fits, it runs smoothly.
Capacity vs bandwidth
Capacity is the size of the bucket. Bandwidth is how fast the GPU can drink from it. A card needs enough of both.
Bandwidth comes from two things: the memory bus width (how many lanes connect the GPU to its memory, in bits) and the speed per pin (in gigabits per second). The formula is simple:
Bandwidth (GB/s) = bus width × speed ÷ 8
So an RTX 5080's 256-bit bus at 30 Gbps gives 256 × 30 ÷ 8 = 960 GB/s.
Image: Zerocade
Notice that the RTX 5060 Ti comes in 8GB and 16GB versions with the same 128-bit bus and the same 448 GB/s. The chip is identical. Only the amount of memory changes, which makes it the cleanest test of what VRAM capacity alone does.
GDDR6 vs GDDR7
Graphics cards use GDDR memory, built for bandwidth rather than the low latency that system RAM prioritises.
- GDDR6 tops out around 20 Gbps per pin. AMD's Radeon RX 9000 cards still use it.
- GDDR7, used across Nvidia's RTX 50 series, launched at 28 to 32 Gbps. Micron says it delivers up to 60% more bandwidth than GDDR6 and over 50% better power efficiency.
The trick behind GDDR7 is how it sends data. GDDR6 uses two voltage levels to send one bit at a time. GDDR7 uses PAM3 signalling, with three voltage levels, which packs 1.5 times as much data into each cycle.
Image: Micron
Faster memory can partly make up for a narrow bus, which is how Nvidia gets away with 128-bit on its RTX 5060 cards. It can't make up for running out of capacity.
What happens when you run out
When a scene needs more than the card has, the overflow goes to system RAM over the PCIe slot. That link is many times slower than VRAM, so the GPU stalls while it waits.
Games handle it in different ways. Some stutter, with frame times spiking even when the average frame rate looks fine; that shows up in the 1% lows, the slowest frames. Some quietly load lower-quality textures or leave them blurry. Some simply crash.
Image: Zerocade
Testing by Hardware Unboxed and TechSpot on the 8GB and 16GB RTX 5060 Ti shows how big the gap gets:
- The Last of Us Part II, 1080p Very High: 109 fps on 16GB vs 67 fps on 8GB. TechSpot measured 1% lows 320% better on 16GB, because the game uses over 9GB.
- Final Fantasy XVI, 1440p Ultra with DLSS Quality: 74 fps vs 41 fps.
- Hogwarts Legacy, 1440p High with ray tracing: the 16GB card had 62% higher averages and 1% lows 483% better. TechSpot called the 8GB result an "unplayable stuttering mess".
- Indiana Jones and the Great Circle crashed on the 8GB card at settings above Medium.
At lower settings the two cards often tie. Alan Wake 2, God of War Ragnarök and Black Myth: Wukong ran roughly even in TechSpot's tests. 8GB isn't broken; it just forces you to turn textures and ray tracing down sooner, even on a GPU fast enough to handle them.
How consoles do it
Consoles skip the split between system RAM and VRAM. They have one shared pool that the CPU and GPU both use, which is part of why games built for PS5 can ask for a lot of graphics memory once they're ported to PC.
Image: Zerocade
The PS5 has 16GB of GDDR6 at 448 GB/s. After the system takes its share, developers get roughly 12 to 13GB, and much of that can go to graphics. A PC port aimed at PS5-level textures is easier to run on a 12GB or 16GB card than on an 8GB one. That's the main reason 8GB has aged so quickly since 2023.
How much do you need?
These are starting points for new games in 2026, not hard rules:
| Resolution | Minimum | Comfortable |
|---|---|---|
| 1080p | 8GB, with textures on Medium or High | 12GB |
| 1440p | 12GB | 16GB |
| 4K | 16GB | 16GB or more |
| 1440p or 4K with heavy ray tracing and frame generation | 16GB | 16GB or more |
If you're choosing between two versions of the same card, like the 8GB and 16GB RTX 5060 Ti, the extra memory is the better buy for anything you'll keep a few years.
Why VRAM is getting expensive
Memory is the reason graphics cards cost more in late 2026. Memory makers are prioritising high-bandwidth memory for AI data centres, and GDDR7 supply for gaming cards has tightened. It's a big part of why the RTX 5080 now costs nearly $2,000 in the US, around ₹2 lakh in India, and why 16GB cards are going up faster than 8GB ones.
When you're shopping, look past the GPU name. Check the VRAM, check the bus, and match both to the resolution you actually play at.
