There’s been lots of other dialogue about the differentiation between integrated analog functions and stand-alone functions. I think one of the conclusions has been that if you need to really squeeze the highest possible performance out of a circuit, it’s probably easier and cheaper to do it with stand-alone analog function blocks. Another conclusion has been that in some cases, integration is the only way to meet the cost objectives, but it carries a lot of risk.
I saw an interesting parallel in some recent market analysis by DSP market watcher Will Strauss. When Will first started tracking the DSP market, the top suppliers were TI, Motorola Semiconductor (now Freescale), AT&T Microelectronics (which became Lucent, then Agere, now part of LSI), NEC, and Analog Devices. All of these companies sold stand-alone programmable DSP chips that were differentiated on speed, power, and to some extent on the instruction set’s ability to execute various algorithms.
All of these suppliers got their chips designed into the digital cellular handset market in the early days. The main function was the voice coder/decoder algorithm, which required lots of multiply/accumulate operations, and could not be done with a general-purpose processor. As inevitably happens in high-growth markets, the race to lower cost and power led to higher and higher levels of integration.
Suppliers focused on that market developed ASSPs (application-specific standard product) with one or more DSP cores embedded (along with a microcontroller, hard-wired logic, and tons of memory), and have displaced the original suppliers at virtually all of the cellphone manufacturers. Strauss now observes that the top supplier of DSP is… Qualcomm. Others in the top tier include Broadcom and Mediatek. And lots of companies now offer DSP cores as IP blocks available for integration into whatever VLSI chip you want to develop. So it appears that the market for stand-alone DSP chips is over. Except that it isn’t.
There is still room for stand-alone DSPs at the high-performance end of the market, like cellular base stations, where software-based modulation and demodulation for emerging standards really push programmable DSPs to the limit. The same is also true at the low-to-medium volume segments, where the risk and development time and cost of a highly-integrated DSP-based digital chip argue for off-the-shelf solutions. In some of these markets, like instrumentation, industrial, and military systems, the benefits of large-scale integration with the DSP function embedded are not as critical as other system features such as the analog performance.
Hmmm. Maybe DSP is more like analog than hard-core analog guys would like to admit. Just don’t ask them to write software.