The PowerPC Has Still Got It (Llama on G4 Laptop) (www.hackster.io)

Vintage computing enthusiast Andrew Rossignol demonstrated that a 2005 PowerBook G4 (1.5 GHz, 32-bit PowerPC, 1 GB RAM) can run a modern LLM by porting a fork of the open-source llama2.c inference engine to the old architecture. He chose the TinyStories model (110M parameters) and had to convert model checkpoints and tokenizer files to the PowerPC’s big-endian format and work around strict memory-alignment constraints by manually copying weights instead of memory-mapping them. Performance was extremely limited — about 0.77 tokens/sec on the stock build (four minutes to generate a short paragraph) versus 6.91 tokens/sec on an Intel Xeon Silver 4216 — but a rewrite of the core matrix-multiply using AltiVec vector instructions nudged throughput to 0.88 tokens/sec. The project is significant because it highlights how software engineering and low-level, hardware-aware optimizations can extend the reach of LLM inference to far more constrained and heterogeneous platforms than we usually consider. It underscores two practical lessons for the AI/ML community: endianness, alignment, and memory-mapping semantics matter for portability, and SIMD/vectorization can yield important gains even where specialized accelerators are absent. While not a practical deployment path, the work showcases the portability of lightweight models and inference engines and encourages continued focus on small-model efficiency and hardware-specific optimizations for edge and legacy systems.