High-end DSP markets compute to higher revenue

Digital Signal Processing (DSP) is a technology that has become pervasive and touches our lives daily. It’s used in our multimedia world of MP3s, DVD players, and digital TVs, to our connected world of cable, DS,L and Wi-Fi modems. These technologies give us access to the Internet, and in cell phones that provide unprecedented mobility, allowing us to reach out and touch someone anywhere, anytime. From a DSP standpoint, many of these applications don't require very much processor horsepower. MP3 decoding, for example, seldom requires a dedicated DSP chip. Just about any low-end RISC chip with a hardware multiplier can do the job.

But, there are DSP applications that require an order of magnitude greater processor capability. Instead of the 50 MMACS (million multiply-accumulates per second) or so required for MP3 decode, there are applications that require several billion MACS or more. The highest-performing DSP chips now available are rated at over 8 GMACS for a single core. Although that's fast enough to address a significant number of high-performance applications – there are even higher-end markets that require multiples of those fast DSPs combined with FPGAs, even multiple instances of both. FPGAs capable of over 100 (18-bit) GMACS are available, so their high-end DSP market presence is assured.

And what may be characterized as a hybrid approach combining FPGA-centric interconnection techniques with collections of processor cores (instead of logic gates), new massively-parallel arrays of elemental processors are now coming to market to serve DSP applications. With as much as 60 GMACS capability and real-time logic reconfigurability, some of these processor arrays are targeting software-defined radio (SDR) applications while others target professional multi-channel, multi-standard video codecs.

Obvious examples of high-end DSP applications include volumetric medical imaging, like MRIs and CT scanners. Early versions of these products were able to scan and later display 3D images over a matter of hours. Newer ones do it in a matter of minutes, and both doctors and patients want them in seconds. All of these scanners require an array of processors to perform the faster processing. High-end DSP chips have also enabled ultrasound scanning in doctor's offices, eliminating the inconvenience of hospital trips for many patients.

Other forms of high-performance scanning include military radar and sonar. These applications can involve dozens of high-performance DSP chips or MPUs with DSP-centric vector coprocessors. And that's in addition to the beamforming DSPs required for electronically-steerable antennas and transducers.

In our latest survey of DSP professionals, we asked participants, "What is the typical DSP performance requirement (in 16-bit MMACS) for your organization's primary DSP-centric product?" The results indicated that there is still a very healthy market for low-end DSPs rated at under 100 MMACS (see Table 1 and Figure 1), but we noted a comparable number of respondents indicated that more than 10,000 MMACS (10 GMACS) were required for their high-end application. Cross-tabulation with answers from other questions indicated that most of those high-end responses involved multiple chips, with as many as 6 chips cited among our survey participants, some with multiple DSPs, others with stand-alone FPGAs and some with combinations of the two. However, several responses indicated single-chip implementations of massively-parallel processors.

  • Medical Imaging
  • HDTV Studio Encoding
  • Video Transcoding
  • Video Encryption
  • Cellular Base Stations
  • Media Gateways
  • Radar & Sonar
  • Explosives Detection
  • Electronics Intelligence
  • Satellite Video Transmitters
  • Wafer/Reticule Inspection

Table 1

Figure 1

Since DSP algorithms are among the most parallelizable (compared with data processing), arrays of parallel processors can make a lot of sense for high-end applications. The key to a successful product of this type is simply the software development and support tools. Graphic algorithm development tools like MatLab and LabView as well as C-language algorithm libraries are common denominators in the DSP world. The ability to map these into new parallel processor array tools can determine the ultimate market success of such products.

In short, if you can't easily program it, the market won't accept it. The poster child for product non-acceptance would be TI's TMS320C80 (a.k.a. MVP) of the last decade. With multiple RISC and DSP engines, it was the most powerful DSP chip on the planet...for the three people who could program it. TI learned the lesson well and its new DaVinci™ multi-core architecture for multimedia applications has become a resounding success.

One final point. The average price of a DSP chip today is $5.70, an average heavily weighted by the 900 million or so DSP basebands shipping in cellphones this year. High-end DSPs, on the other hand, sell for hundreds of dollars, while high-end FPGAs can cost several thousand dollars. Consequently, the high-end DSP market is best understood in total revenue, not units. Of this year's predicted $25 billion DSP-centric chip market, Forward Concepts sees the high-end DSP market as a $2-billion opportunity for chip companies that can fulfill industry's needs.