Replace the M5 Max's LPDDR5X with eight Samsung LPDDR5X-PIM packages and the memory could deliver 4.9 TB/s where the weights live. The right parallel mapping could push Llama 3.1 8B toward 650 tokens per second, or a 27B dense model toward 190.
Samsung's LPDDR5X-PIM puts compute inside commodity memory, exposing 614 GB/s of bank-level bandwidth for local LLM inference. The silicon works. Quantization, memory layout, and runtime support now stand in the way.
How I built real-time 3D rendering on a slow 486 with only 256 colors in VGA Mode 13h, using fixed-point math, palette quantization, matcap-style Phong lighting, and x86 assembly.
Today, I built a pure-TypeScript Zstandard implementation (library + CLI + dictionary training + interop tests) in just 4 hours using OpenAI Codex 5.3.
