Modern Computer Architecture and Organization — From the Processor to the AI Data Center (Books)
I have the first edition of Jim Ledin's Modern Computer Architecture and Organization on my desk and decided to use it to refresh my knowledge of computer architecture. Then I learned that a third edition is already out, with dedicated chapters on GPUs and LLMs. GOTO Book Club has a conversation with the author that explains well how the AI boom reached this book.
The book itself is a step-by-step journey through how a computer works, from the bottom up. It starts with transistors, digital logic and processor components, then moves to memory, input/output and the hardware–software interface. Next come caches, pipelines, parallel execution, x86/x64, ARM and RISC-V. It also reaches the system level: virtualization, specialized computing, and the design of smartphones, PCs and cloud servers. The first edition was published in 2020.
For developers, the value lies in how the pieces connect: from the instructions a processor can execute to how it fetches data and runs a program. You can rebuild the picture beneath familiar abstractions and understand why performance depends on far more than CPU frequency. This sequence works well for revisiting the fundamentals.
While looking for reviews of the book, I noticed that the third edition came out on March 31, 2026. The first edition already covered GPUs, neural networks and data-center-scale computing, but the third gives GPUs and LLM compute architectures their own chapters. In the interview, Ledin explains how he wrote about a field that changes faster than a publishing cycle: he focused on enduring principles — parallelism, memory bandwidth, tensor operations and the connections between computing nodes.
He chose GPT-2 to explain LLMs. According to Ledin, its open code and relatively small size make it possible to trace how a transformer works: which matrices are multiplied, how the layers are arranged and what happens to the data. You can work through and experiment with that teaching example yourself. Ledin believes understanding GPT-2 gives you a good foundation for approaching larger models. That is a sensible way to teach architecture: understand the mechanism first, then increase the complexity.
The conversation then turns to how this design grows from a single chip into a server, a rack and an entire data center. Ledin points to memory bandwidth as one of the key constraints: compute units can sit idle while data is read from memory or written back. At a larger scale, communication between devices, power and cooling add further constraints. He suggests thinking of the whole data center as one large computer, with its own architecture and limits.
Toward the end of the interview, he is asked which single chapter he would recommend to a developer without a hardware background. Ledin chooses the chapter on improving performance: pipelines and the cache hierarchy. Different instructions are processed simultaneously at different pipeline stages, while caches keep the needed data closer to the core. In his view, understanding these mechanisms helps us organize code and data so the hardware spends less time idle. We made it all the way to AI, but still have to revisit the basics :)
The first edition I already own remains useful for revisiting processors, memory and program execution. If you want to continue that journey into GPUs and systems for LLMs, the third is worth a look. And the interview is a good way to understand the thinking behind the update before taking on the new chapters.
#Books #Engineering #Architecture #Hardware #AI #SoftwareDesign