Remarkable Throwback: How a 1997 Processor and Just 128 MB of RAM Can Power Modern AI—Discover the Proof!

In a surprising twist, researchers from EXO Labs are showing that powerful AI doesn’t always need the latest hardware. Instead, they ran a modern AI model on a computer from 1997—a Pentium II.

The adventure began when the team found an old Pentium II on eBay for £118.88. It had just 128 MB of RAM, which is tiny compared to today’s standards. But EXO Labs was eager to explore the possibilities.

Setting up the machine was easier said than done. With no USB ports, they had to use PS/2 for peripherals. Funny enough, the mouse had to go in port 1 and the keyboard in port 2. This small detail nearly derailed the project!

To transfer files, they faced another challenge. USB drives couldn’t be used, so they set up a network between the old PC and a MacBook via FTP. Using an adapter, they overcame this hurdle and began transferring essential files, like model weights and codes.

Next, they tackled the task of compiling modern code for such an old machine. They initially tried some newer tools but hit a wall since the Pentium II couldn’t process them. Eventually, they turned to Borland C++ 5.02, which still worked perfectly.

After tweaking the code, they ran the Llama 2 AI model. It was a groundbreaking moment: the 260K parameter model processed 39.31 tokens per second! While it’s slower than today’s AI, running it on hardware this old was impressive. They attributed their success to the BitNet architecture, which simplifies calculations.

This experiment reveals a new possibility in AI. Instead of always needing the newest tech, could we make AI accessible on older hardware? The EXO team is already working on BitNet, designed to be efficient on older machines. For instance, a large BitNet model might require only 1.38 GB of storage, making it reachable for older laptops and even some gaming consoles.

This could potentially democratize AI, giving more people access to its capabilities. Whether through old devices or low-powered systems, we might soon see AI integrated into a broader array of technologies.

As we ponder the future of AI, this experiment reminds us that sometimes, less truly can be more. By focusing on efficiency, we can bridge the gap between old technology and new demands in the tech landscape.



Source link