Jeff Bier’s Impulse Response—Caveat Emptor

Submitted by Jeff Bier on Tue, 05/01/2001 - 18:00

In the ancient Near East the number 40 was a literary device symbolizing "an impressively large number," not a precise quantity. "It rained for 40 days and 40 nights," "they wandered in the desert for 40 years,"—the examples are countless.

In the modern world of DSP processors, the term "10X" has taken on a similar function. Vendor claims that a processor is, e.g., "10X faster than existing DSPs," often don't mean that the processor is really 10 times faster—only that it is "much faster." The problem, I suppose, is that "much faster" just doesn't resonate as strikingly in a press release.

While overblown marketing claims are nothing new, the rapid expansion of the DSP market in recent years has created an increasingly competitive sales environment. This has made processor vendors—even some of the very largest—desperate to get our attention. Many have resorted to hyperbole, which can obscure a new processor's true capabilities.

In the old days (say, five years ago), DSPs generally had very similar architectures. This made comparing processors fairly simple: the main performance differentiator was clock speed, and simple performance metrics like MIPS provided useful information.

The increasing diversity of both DSP architectures and applications has now rendered clock speed and simple performance metrics useless in assessing processor performance, and has complicated processor comparisons. We frequently see cases where processor A is twice as fast as processor B on, e.g., an FFT, but twice as slow on another DSP algorithm, e.g., a Viterbi decoder.

This highlights the importance of using appropriate benchmarks to compare processors for DSP applications: not just any benchmarks, but benchmarks that are representative of the target application.

Unfortunately, hyperbolic performance claims ultimately hurt everyone involved. System designers who naively accept exaggerated claims may find—often too late—that their systems don't meet their design goals. Reporters and market analysts who parrot these claims lose credibility with their readers. Processor vendors making such claims lose the trust of their customers, if not their business.

These days, getting educated is the only way to avoid getting burned: processor performance claims should be evaluated with both caution and a healthy dose of skepticism.

Add new comment

Log in or register to post comments