Modern Computer Architecture and Organization

Jim Ledin — author of Modern Computer Architecture and Organization, now in its third edition — joins Maciej "MJ" Jedrzejewski to talk through the book's two brand-new chapters on GPUs and large language models, and the specific challenge of writing about hardware that changes faster than a book can ship. Ledin's approach was to focus on durable architectural principles (parallelism, memory bandwidth, tensor processing, and how systems scale from one chip to an entire data center rack) rather than chasing the current state of the art. For the LLM chapter specifically, he chose GPT-2 as the teaching example precisely because it's open-source and small enough to trace by hand — and because, as he explains, the transformer architecture underneath today's frontier models is fundamentally the same structure, just scaled up.

The most concrete takeaway for engineers: the AI buildout's real bottleneck right now isn't compute, it's memory bandwidth — a shift driving high-bandwidth memory production at the expense of consumer RAM, which Ledin illustrates with a blunt data point (a 32GB DDR5 stick that cost $69 a year ago now runs $440). But asked which single chapter would most change how a working developer with no hardware background thinks about their own code, Ledin doesn't point to AI at all — he points to the chapter on pipelining and cache hierarchy, arguing that most of the thousandfold speedup from 1980s PCs to today came from those optimizations, not clock speed, and that writing cache- and pipeline-friendly code is still one of the highest-leverage skills a developer can have.