GPU stands for Graphics Processing Unit. It was built to draw game graphics, but AI work — huge numbers of multiplications and additions — has the same shape. A CPU has a few powerful cores; a GPU has thousands of small ones working in parallel.
That is why a model that takes weeks to train on a CPU can finish in hours on a good GPU. Data-centre GPUs such as the NVIDIA L4, A100 and H100 are designed specifically for AI.
Llama 3.1 8B writes about 2–3 words a second on a laptop CPU, and roughly 30–40 on an L4 GPU (4-bit).
বাংলায়: GPU (জিপিইউ) — একসাথে হাজারো ছোট হিসাব করতে পারে এমন প্রসেসর — AI চালানো ও শেখানোর মূল যন্ত্র।