Parallel Computer Organization and Design

The field of computer architecture and microprocessor design has undergone a significant change in the past decade. The three best books about computer architecture that I know of are: Computer Architecture, Fifth Edition: A Quantitative Approach Modern Processor Design: Fundamentals of Superscalar Processors Parallel Computer Organization and Design  (this book)Of these three books, Parallel Computer Organization and Design has the best coverage of the issues that have limited the increase in single core performance, as well as important constraints in the development of multicore processors.Despite the title, this book covers all the important elements associated with the design of modern high performance CPUs, not just parallel computers.The first four chapters focus on topics that don't directly deal with parallelism, but matter more for the design of the components in a CPU. These chapters help motivate the demand for parallel computers by highlighting limitations of scalar architectures, but they also give a good overview of fundamental concepts for the design of a CPU core. Chapters 5 through 8 focus on topics related to designing a parallel system, while chapter 9 provides a survey of simulation techniques used in the development of modern CPUs.The book is organized in a logical fashion and it is instructive to read it from cover to cover. However, each chapter also provides standalone coverage of a specific topic and can be read individually. The chapters are also comprehensive and the book does not assume that the reader has an extensive background in computer architecture. I think it would be useful as a textbook for someone in a computer architecture course, or simply as a good reference on modern CPU design.My primary complaint with the book is the lack of references. There are none. While the authors clearly drew on a vast amount of personal knowledge, mistakes or omissions are inevitable when covering so much material. For instance, I noticed that the explanation of the SMARTS sampling technique on pages 512-513 was not completely correct. This made me somewhat less confident about other topics explained in the book that I don't have as much expertise in. Having references would have added to the credibility of the book.Another failing is that the book does not differentiate clearly between topics that are implemented in commercially available CPUs and ones that currently only exist as research proposals. This was most notable in the explanation of thread-level speculation in section 8.5.4. Also, Chapter 9 did not differentiate between simulation techniques used primarily in academic research and those used for development of real CPUs.Overall, however, I found the book to be a clear and comprehensive survey of the most important topics related to the design of modern CPUs, and I highly recommend it for anyone wanting to learn and understand the details of how modern CPUs work.

good

Good

The book is well suited as a text for a graduate level course in parallel architecture. The authors start with a basic device physics description of the silicon. Describing fundamental properties of the chips, and failure modes like electromigration along a conducting line in the silicon. Power minimisation is broken into 2 parts - for static power and dynamic (AC) power. The leakage current of a transistor in the off mode is often the limiting factor in determining the minimum power loss.Another chapter then goes into the different types of processor architecture currently built. Here concepts like how to use or rather, maximise, pipelining and the efficacy of branch prediction are discussed. The latter is interesting, in that the text explains how it is implemented in the hardware, and thus outside the control of any software you might load into the CPU.The technique of Very Long Instruction Word architecture has been around for decades. As we see here, it is still occasionally used. Indeed, the text describes several VLIW implementations.Another key idea is the use of memory hierarchies. Fast memory caches can or should sit on the same chip as the CPU. The cost of these is simply the cost of the real estate area on the chip. You cannot have a lot of fast caches. But nowadays it is common to have L1, L2 and L3 caches, in order of increasing access time. While off the CPU, there sits main memory, which should be familiar to you, because these are the explicit (if I may use this word) memory chips that are quoted in size whenever you look at the specs of a new computer. Main memory is now several gigabytes, but the problem is that accessing it takes around 200 ns, while getting at the slowed L3 cache is only 20 ns.The book's strength is that it explains these and other concepts clearly and simply. The narrative gives enough detail for you to appreciate and understand the ideas, without introducing unnecessary clutter.

This book features a wide range of parallel computer architecture topics, from introductory to advanced. The importance of parallel computer architecture is motiveted in the beginning and the authors give a quick but still quite detailed overview of what the reader should already be familiar with. It is well-structured and easy to read. Additionally, the chapters are mostly independent of each other.I would definitly recommend this book to anyone with prior knowledge of computer architecture.

This is the book I wish I'd had when I started my MSc by research 2 years ago. It is very well written highly detailed and doesn't spare on the technicalities, but without getting too abstract and mathematical. While the field of parallel computing hasn't changed dramatically in the last 10-15 years there are a lot of useful updates as well as a lot of well established theory included for completeness. If you only read one book on parallel computing, make sure it's this one.