Guided pathThis is part of Understand Transformers & BERTBack to the path

Why LLMs Live in Data Centres

P3.computing-fundamentals.03 · Audience: guest, it-ml, language-pro · Prerequisites: GPUs & Parallel Computing

Real LLM grading for this pageLLM grading (this page):

You can use a powerful AI tool on your phone, but you cannot train one on your desktop. Why? The answer is arithmetic you can do yourself: a model's size in bytes, the memory a GPU has to spare, and the extra memory training demands on top of that. This module turns abstract "billions of parameters" into concrete gigabytes — then lets you compute the fit for yourself in two short exercises.

Step 1 / 3How big is a language model?

A language model is a pile of parameters (also called weights) — the numbers learned during training. Each is usually a 32-bit floating-point number taking 4 bytes. So the size is simply:

model size (bytes) = number of parameters × bytes per parameter

ModelParametersSize (FP32)Fits a consumer laptop?
BERT-base110 million~440 MBYes (16 GB)
GPT-2 large774 million~3 GBYes
Llama 3 8B8 billion~32 GBBarely (32 GB, tight)
GPT-3175 billion~700 GBNo (needs ~10 GPUs)
GPT-4 (est.)>1 trillion>4,000 GBNo (a data centre)
Size in GB (log scale) — each step is roughly x10BERT-base440 MB · fits a laptopBERT-large1 GB · fits a laptopGPT-2 large3 GB · fits a laptopLlama 3 8B31 GB · tight (32 GB)GPT-3684 GB · data centreGPT-4 (est.)3906 GB · data centre
Model size = parameters x 4 bytes (FP32), on a log scale. Green fits a consumer machine, amber is tight, red needs a data centre.
🗣️ From a linguist's perspective: Model size as a printed reference library
A single dictionary entry is roughly 200 bytes of text; a full bilingual dictionary of 100,000 entries is about 20 MB — a USB key. BERT-base stores 110 million numbers, about 5,500 such dictionaries. GPT-3's 175 billion parameters would fill roughly 8.75 million dictionaries. The trouble with running GPT-3 on a laptop is not software — it is physical: you cannot put 8.75 million dictionaries on a kitchen table.
Quick recap — model scale
  • Model size in bytes = parameters × 4 (FP32). GPT-3 is ~700 GB; BERT-base is ~440 MB.
  • VRAM is the GPU's own memory — the whole model must fit there. Training needs ~4× the model's weight size (weights + gradients + optimiser state).
  • Inference (using a model) is cheap and fits consumer hardware for small-to-mid models; training frontier models needs thousands of GPUs and months of compute.

Practice — Ladder E: does it fit?

Two short exercises turn the arithmetic of this module into code. The first sizes a model in gigabytes; the second decides whether a model fits a given GPU, for both inference and training. Both run in your browser — no setup, no data centre required.

Ask the mentor about this module

Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.

Ctrl/Cmd + Enter to send

Loading exercise…

Loading exercise…

📚 Go Further

Where the numbers behind model scale come from.

TypeResource
Video3Blue1Brown — How large language models work (YouTube)
ReferenceTim Dettmers — Which GPU(s) to Get for Deep Learning (timdettmers.com)
ReferenceEpoch AI — Trends in AI (epochai.org) — compute, dataset size, and training cost over time