Serving technology enthusiasts for more than 25 years. TechSpot is the place to go for tech advice and analysis you can trust.
Hot potato: Open-source project llama2.c was designed to run a lightweight Llama 2 version entirely in C code. This “baby” Llama 2 Model is inspired by llama.cpp – a project designed to enable LLM inference on a wide range hardware, including local devices and cloud-based platforms. These compact code experiments can now be used to run AI technology on any device with a processor, highlighting the increasing accessibility and versatility of AI.
After watching Exo Labs run a large Language Model on an ancient Pentium 2 running Windows 98 on a Pentium II, developer Andrei David decided on tackling an even more unconventional task. He dusted off his Xbox 360 and set out to force this nearly two-decade old machine to load a Meta AI Llama LLM.
David announced on X that he had successfully ported Microsoft’s gaming console from 2005. The process was not without its challenges. The Xbox 360’s PowerPC processor is a Big-endian required extensive endianness translation for both the model’s weights and configuration. He also had to make significant adjustments and optimizations in order to get the code to work on the old hardware.
LLM on Xbox 360
Xenon CPU (3GHz PowerPC w/ three cores) with 512MB unified memory. Based on ahref=””https://twitter.com/karpathy?ref_src=twsrc%5Etfw””>@karpathy ( “sllama2.c”ported to run in Microsoft’s powerful console since 2005. Pure C implementation optimized to run on PowerPC architecture with Xbox memory management.
Inspired by… pic.twitter.com/e9oMLaWIyi
– Andrei David (@AndreiDavid) January 10, 2025
Memory management posed yet another significant challenge. The 60MB llama2 had to be carefully crafted to fit into the Xbox 360’s unified architecture of memory, where CPU and GPU share a pool of RAM. David believes that the Xbox 360’s memory design was very forward-thinking at the time. It foreshadowed the memory management techniques used in modern gaming consoles.
David ran llama2 successfully on his Xbox 360 after extensive coding and optimizing.Despite being only 700 lines of C code without external dependencies, David noted it can deliver strong performance “surprisingly” when tailored to a narrow domain.
Since the developer accepted the challenge, we can expect to see more LLM experiments on Xbox 360 soon.
