The author attempts to run matrix multiplication code in Swift matching C's efficiency for LLM training. Initial Swift performance lagged at 2.8 Gflop/s, producing 1 token every 19 seconds. This stark contrast to C's 1.911×10¹¹ FLOPs highlights Swift's optimization gaps. The challenge stems from Swift's higher-level abstractions and safety checks, which introduce overhead compared to C's raw execution. The article documents incremental optimizations, including manual memory management and loop restructuring, to bridge this divide.

Key insights emerge from instruments profiling. The biggest bottleneck was _ArrayBuffer.beginCOWMutation(), consuming significant CPU cycles. By eliminating bounds checking and leveraging Swift's release mode, performance improved but remained subpar. The author experiments with Span and Egg data structures to reduce memory overhead. These tweaks reflect a hands-on approach to understanding Apple Silicon's hardware capabilities, particularly the AMX unit designed for matrix operations. The work underscores the gap between Swift's developer-friendly design and C's computational precision.

The project's significance lies in its practical implications for Mac-based LLM development. While frameworks like PyTorch abstract hardware, this manual approach reveals how Apple Silicon's architecture—especially SIMD and AMX—can be harnessed. The author's goal isn't just speed but transparency: showing how low-level optimizations impact real-world ML workloads. By comparing Swift to Karpathy's C baseline, the work serves as both a technical deep dive and a call for better Swift tooling in machine learning.