<- All terms

VRAM

A GPU's dedicated memory. It has to hold the model weights, the KV cache for every in-flight request, and activations. When it runs out, you need a bigger GPU, more GPUs, or a smaller model.

What it is

A rough weights budget: ~2 bytes per parameter at 16-bit precision. A 70B model is ~140 GB in fp16, ~70 GB at int8, ~35 GB at int4 - see quantization. On top sits the kv cache, which grows with context and concurrency.

Why it matters

VRAM, not compute, is usually the first wall you hit when self-hosting. It decides which single GPU can run a model at all, and how many users it can serve before you need multi gpu parallelism.

Cost & infrastructure impact

Every VRAM tier up roughly doubles the hourly price (24 GB -> 48 GB -> 80 GB -> 141 GB). quantization is the main way to fit a model into a cheaper tier, at some quality cost.

Example

Model, precisionWeights (approx)
8B, fp16~16 GB
70B, fp16~140 GB
70B, int4~35 GB

Related concepts

Use it in Obolith

FAQ

How much VRAM do I need to run a 7B model?

About 16 GB in fp16 for weights plus headroom for KV cache - a 24 GB card is comfortable. At 4-bit you can fit it in ~6-8 GB.

Last reviewed: 2026-09-01