1. Well, not really.2. But I like to think of myself as a reasonable person, so I'm sensitive to the fact that it is unhealthy for a community to become known for initially astonishing claims that - maybe after some deeper digging or a leavening of experience - don't quite hold up under careful scrutiny.

3. The natural successor to the wave of "GPU Hype" is the "Downer of Realism" where the afterglow of stunning performance reports is angrily dismissed; like a sunday school teacher breaking up a game of truth-or-dare in a back room on church youth night.

(err...that's a totally hypothetical example!)

4. As cases in point, consider this paper from Intel

or

This blog post.

5. I have a hard time rebutting the notion that the speedups I've realized by employing GPU-accelerated code would seem less spectacular if, instead of my own vectorized Matlab code as a baseline, I had a team of Intel engineers convert my original programs to multi-threaded C++ or FORTRAN with hand-tuned cache optimization and memory mappings on top-of-the-line hyperthreaded processors.

6. But still, I'll admit to being happy when, after a short time of learning CUDA and building upon what I admit is a fairly broad yet unspectacular programming knowledge base, the sort of Matlab program that an ordinary engineer (like me) can write is converted into a CUDA program that an ordinary engineer (like me) can produce - except that it happens to run more than 100x faster than the original version.

7. Yes: as part of an alternate strategy I converted my Matlab to C++ (with a decent compiler); and
Yes: the C++ was only on the order of 15x slower than the CUDA/C++ version; and
No: that C++ code did not exploit multi-threading; and
No: my CPU is not top of the line. (of course, neither is my GPU....but anyway...)

8. My points for the individual engineer/programmer are:

a. The speed-up that matters to you, is the speed-up that you can get with the equipment and programming skills you have.

I'm sorry that Intel sometimes doesn't seem to understand that. If they did, IMKL would be much easier to use, nobody would bother with Matlab or CUDA or Jacket or libJacket and we would be talking about how great Intel's parallel programming model is, not those of some other guys.

Source: noiceinmyscotchplease.blogspot.com