Skip to content
Bitpute
ToolsModelsGPUsCloudLearn
AI HARDWARE INTELLIGENCE

Know What
Your AI Needs.

Calculate VRAM, check GPU fit, compare hardware and plan AI workloads before you buy or rent.

Calculate GPU Needs Explore Tools
Used by 200K+ builders, researchers
and AI teams worldwide.
Model
Llama 3.1 70B ›
Parameters
70B
Precision
BF16
Context Length
32K
VRAM Required 146.2 GB
GPU Fit Score
92%
Excellent
NVIDIA H100 80GB
Live GPU Database
583+ GPUs
TRUSTED DATA & METHODOLOGY

Three numbers
decide the fit.

The same engine runs on every page — no vendor spin, no crowd-sourced guesses.

01

Weights params × bits ÷ 8

The model itself. A 70B model at Q4_K_M is about 40 GB before anything else touches memory.

02

KV cache grows with context

Every token you feed the model is cached. At 128K context, the cache alone can rival the weights.

03

Overhead CUDA + fragmentation

A fixed floor of roughly 0.75 GB plus a small fraction of the weights, reserved before your first token.

Parity-tested against the JavaScript engine to the gigabyte. See the full methodology

See it at scale

Memory bends with context.

Llama 3.3 70B at Q4_K_M. Total VRAM climbs as the context window grows — the number a spec sheet never shows you.

Context lengthTotal VRAM · GB

Two ways in.

For builders

The Workspace

Type a model, pick a budget, and get the fit, the decode speed, and the buy-vs-rent call ranked across every card — on one screen.

Open the Workspace
For the curious

The Engineering Center

Every formula, constant, and source behind the numbers — the methodology laid bare, with the reasoning you can check line by line.

Read the engineering

Questions, answered plainly.

How much VRAM do I actually need?
Enough for weights, KV cache, and a little overhead. A 7B model at Q4 fits in 8 GB; a 70B at Q4 needs roughly 48 GB before context. Set your model and card in the engine above for an exact figure.
Is a card with more memory always better?
Only up to what your model needs. Past that point, memory bandwidth — not capacity — sets your speed. The engine reports both, so you can see which one is actually limiting you.
My model is bigger than my card. Can I still run it?
Usually. Quantize it to fewer bits, offload some layers to system RAM, or split across two GPUs. Each path trades memory for speed, and the calculators show that trade in tokens per second.
Should I buy a GPU or rent one online?
It comes down to hours per month. The Workspace computes the break-even from each card's purchase price and its hourly cloud rate, so the answer is a number, not a hunch.
Why are your numbers higher than the model's own page?
Model cards usually list the weights only. Real deployment adds KV cache and CUDA overhead — the memory that actually has to fit on the card. Bitpute counts all three.

Stop guessing. Size it right.

Free, private, and cited to the last gigabyte.

Start calculating Compare GPUs