AI Compute RadarVRAM vs. System RAM
VRAM is the memory on your graphics card, system RAM the memory on the motherboard. For local AI the difference is speed, not size: a model must mostly fit into VRAM to run fast.
Is VRAM the same as RAM? No. Both hold data for a processor, but VRAM sits on the graphics card and is wired for the bandwidth of the GPU, while system RAM serves the CPU over a much narrower path. Adding RAM does not add VRAM, and the two cannot be swapped: a model that spills out of VRAM runs at the speed of the slower path.
A PC has two kinds of memory. The graphics card carries its own — VRAM — small but extremely fast; the motherboard carries system RAM, plentiful but far slower for this workload. Generating text touches a large share of the model's weights for every single token, so generation speed is effectively memory speed.
That makes VRAM the number that decides your local AI class. Consumer cards span roughly 8 to 32 GB, and each step up admits a bigger tier of models at comfortable speed. System RAM size barely matters for models that fit in VRAM — and matters a lot the moment they spill over into offloading.
Two footnotes worth knowing: laptop GPUs often carry much less VRAM than desktop cards of the same name, and Apple Silicon plays by different rules entirely with its unified memory.
Where you see it on the radar
Pick your exact GPU on the radar and every tracked model gets a verdict against its real memory — measured sizes, stated margins, no rounding up.
Get verdicts for your card →Related terms