The 16GB Line: GPU VRAM Tiers for Local LLMs, Priced and Counted
VRAM tiers are cliffs, not slopes. A model fits whole on your card or it spills into system RAM and slows to a crawl, and the cliffs sit at the same memory sizes for everyone. We sorted all 171 GPUs in this directory into the seven tiers the market actually sells, priced the cheapest way into each one, and wrote down what each tier honestly runs. The short version: below 16 GB you are in quantization territory, 16 GB is where decent local models start, and the really good ones are waiting at data center prices. For now.
The tiers, counted and priced
| VRAM tier | GPUs in directory | Cheapest way in (US) | What honestly fits |
|---|---|---|---|
| 8 GB or less | 19 | Intel Arc A580, $179 | 7B to 8B models at INT4, and nothing bigger |
| 12 GB | 29 | Intel Arc B570, $219 | 8B at INT8, 14B at INT4 |
| 16 GB | 53 | Intel Arc Pro B50, $299 | 8B at full precision, 14B at INT8, the low 20B class at INT4 |
| 24 GB | 28 | Intel Arc Pro B60, $599 | 32B at INT4 comfortably, 70B with offload |
| 32 GB | 11 | NVIDIA GeForce RTX 5090, $1,999 | 32B at INT8; one card starts to feel roomy |
| 48 GB | 8 | NVIDIA A40, $3,250 | 70B at INT4, whole, no offload |
| 80 GB or more | 23 | AMD Instinct MI210, $5,499 | 70B at full precision, and several models at once |
Below the line: quantization territory
Everything under 16 GB runs small models squeezed hard: aggressive quantization, short context, no headroom. The famous name in this zone is the RTX 3080 Ti, and it sits in the 12 GB tier next to cards a fifth of its price, because the tier asks only one question. Apple is not in this table at all. Unified memory is not a tier, and between the bandwidth ceiling and the missing CUDA ecosystem it is not a serious way to run local models, a case we made in full in our Mac Studio vs RTX workstations comparison.
The 16 GB line, the 3090 tier, and the top
The 16 GB tier is the biggest in the directory, and it is where the models people actually want start to fit whole: 8B at full precision, 14B at INT8, the low 20B class at INT4. The cheapest NVIDIA way in is the GeForce RTX 5060 Ti 16GB at a $429 list price. One tier up, the 24 GB cards remain the used market's gift: the RTX 3090 class runs 32B models at INT4 all day and drags 70B along with offload. At 48 GB a 70B model fits whole at INT4, and at 80 GB and above you are in H100 and H200 country, $27,999 and up in our data, where 70B runs at full precision and several models run at once.
The data center end does not survive
That top tier is the one living on borrowed time. A college student with a good GPU already runs a language model, a voice model and a video model on one desk, in sequence or side by side, for the price of the card. Every model generation squeezes more into 16 and 24 GB, and every card generation puts more memory on a desk for less money. The cloud keeps the work that is genuinely huge; the everyday stack moves home. Some startup will build the machine, or the new device, that makes that the default, and when it does, the shopping list looks like a directory like this one. Until then: buy the tier your models need, check the fit with the VRAM Fit Checker, and read the full VRAM ladder before you spend.