AI Model Runs on Xbox 360: See How 3-Core PowerPC with 512 MB Tackles Llama2.c!

Andrei David, a user of X (formerly known as Twitter), recently managed to install and operate an AI model based on Andrej Karpathy’s llama2.c on an Xbox 360, a gaming console that’s nearly two decades old. Using a bootloader, David was able to run a prompt titled ‘Sleep Joe said’ on the vintage hardware. This experiment was inspired by EXO Lab’s successful operation of the Llama model on a Windows 98 PC powered by a Pentium II processor.

The AI large language model (LLM) utilized by the Xbox 360 is similar to that used by EXO Lab on the Windows 98 system. However, David noted that he had to adapt the code for the console’s PowerPC architecture and its distinctive memory management system that was advanced at its launch. The most noteworthy difference is that PowerPC is a big-endian system, which means it stores the most significant byte first.

In contrast, the Intel Pentium II processor that EXO Lab used is little-endian, storing the least significant byte first. To get the AI model working correctly on the Xbox 360, David implemented a byte-swapping technique to ensure that any data was properly aligned with 128-byte memory segments, as required by the Xbox 360’s memory system.

David also had to navigate the challenges posed by the Xbox 360’s unified memory architecture, which was quite innovative at the time of the console’s release. This setup, now common in modern consoles, required developers to optimize how games utilized the hardware’s capabilities. Although the console featured 512MB of RAM, David had to ensure the 60MB AI model would work within the memory allocated to either the CPU or GPU, as these components shared the same RAM pool.

Through meticulous memory management and adjusting to the PowerPC’s operations, David was able to get the old Xbox 360 to successfully run the LLM. This achievement moves EXO Lab closer to its vision of democratizing AI technology, making it accessible not just to large corporations with deep pockets and high-tech resources, but to everyday users.

Executing a Llama 2-based AI model on such an old device like the Xbox 360 is impressive in itself. One user remarked that the 512MB of unified RAM should suffice for running simpler models like Hugging Face’s SmolLM or the 4-bit 0.5B Qwen2.5 model. To this, Andrei David responded enthusiastically, ‘Challenge accepted! :)’, hinting at future attempts to bring even more robust AI models to the Xbox 360.