How does generative AI work?
Generative models learn from very large amounts of data which patterns are found in text, images or sound. From these patterns, they compute new content that matches the input. Text models such as large language models (LLMs) predict word by word what comes next. Image generators usually work with diffusion models. They start with image noise and refine it step by step. In both cases, the technical foundation is deep learning.
What hardware does generative AI need?
Training large models requires enormous computing power. Research institutions and companies use GPUs for this because they can execute thousands of computing operations simultaneously. For training, operators connect many GPU servers to form an HPC cluster. Day-to-day operation also requires a lot of GPU memory. NVIDIA GPUs such as the H200 and B200 are widely used for this.
You can test in advance which configuration suits your model in the MEGWARE Benchmark Center.
FAQ Generative AI
AI is the umbrella term for systems that solve tasks with human-like intelligence. Generative AI is a subfield of it. It creates new content such as text, images or code instead of merely classifying existing data.
Well-known examples are ChatGPT for text, Stable Diffusion and Midjourney for images, and GitHub Copilot for program code. In research, generative models propose new active ingredients or materials.
Generative models compute billions of similar computing steps, above all matrix multiplications. GPUs execute such calculations in parallel on thousands of cores. That is why data centers today train and run generative AI almost exclusively on GPUs.
That depends on where the data is processed. Many cloud services process inputs on external servers, some of them outside the EU. If you run models in your own data center, you keep full control over your data.
Yes. Open models such as Llama or Mistral can run on your own hardware and be adapted with company data. Which GPU configuration you need depends on the model size and the number of users. You can test this in advance in the MEGWARE Benchmark Center.