This AI language model has 10 times lower memory consumption: Nvidia and AMD are the target

Chatbots like Gemini or ChatGPT use a large language model (LLM) with generative artificial intelligence to perform different actions. The possibilities of these tools are increasing, but so are the hardware requirements of the devices to run them.

A study conducted by the University of California has created a new LLM that can run billions of actions in the background with 13 watts of power and without losing performance. This figure is equivalent to the consumption of lighting a 100 W light bulb.

This was one of the main concerns of users. Windows 11 includes AI features, but many people claim that their device has decreased performance. Some people have even decided to go back to Windows 10.

The key to the success of the University of California project is not investing in more powerful systems, but radically changing the way information is connected. The LLM creates a neural network that runs on custom hardware and eliminates the multiplication factor of many AI systems.

The LLM creates a hierarchy between words

The algorithms of Current LLMs are based on the multiplication of elements based on associations. Words are represented as numbers and generate matrices with which to operate to obtain a numerical result that is equivalent to a linguistic one.

The system assigns a hierarchy to words and relates them to others in a sentence or paragraph. The arrays are stored on hundreds of separate GPUs and retrieved with a new query. This process avoids transferring data that must be multiplied across hundreds of arrays, so energy consumption is lower.

The LLM is simpler, as each word equals a number and each number has a value between negative one, zero or positive one. The systems They just have to add the numbers instead of multiplying themso the algorithm saves time and reduces hardware consumption.

The research team has created custom hardware, but the goal is to convince chip manufacturers like Nvidia or AMD. The software is open source and has already been tested on the first standard GPUs with a 10 times lower memory consumption and a 25% increase in speed.

Known how we work on Computertoday.

Tags: Artificial intelligence, Chip, Processors