Only recommended for die-hard CUDA fans

I had already read many articles from the series “CUDA, Supercomputing for the Masses” by Rob Farber on the Dr. Dobb’s website and was quite impressed. Back then in 2009, it was one of the few ways to read something more sophisticated about CUDA.

The title “CUDA Application Design and Development” of the book gave me the impression that it would cover general guidelines and principles for creating CUDA applications.

Unfortunately, the book is more of a collection of articles on topics that apparently interested the author but have no place in a textbook about CUDA. Machine learning may be a very interesting topic, but then the book should have been called “Machine Learning with CUDA.” The author does not do himself justice here.

Here are my points of criticism in detail:

Ch. 2 is far too preoccupied with the theory of machine learning. Functors can also be used in simpler algorithms.
Macros should not be used in C++ (p. 47).
An #include in the middle of the code is also bad (p. 47).
Ch. 8 about executing CUDA code on x86 processors refers to a commercial product from the company PGI. This is actually a chapter of advertising and not labeled as such.
The 4.5 pages of example code contain a lot of mathematics but little CUDA (p. 57ff).
In Chapter 9, a lot of OpenGL code is printed. This is not of interest in a CUDA book.
The author describes the GPU Ocelot framework as “popular, actively maintained” (p. 187). This is a massive exaggeration; there was only one presentation per year on this topic and only one release annually.
SWAN and MCUDA were once research projects and, according to their websites, have not been active since 2010.

Nevertheless, there are a few interesting parts in the book. However, you don’t need to buy the book specifically for these, as they are explained better in newer books. But if you are an enthusiastic CUDA fan, you can read the book and possibly get a few ideas.

Rob Farber
CUDA Application Design and Development
Morgan Kaufmann
2011