Digital Signal Processors Come of Age

CABLE WORLD STAFF

Looking back at the history of programmable chips - in a lecture available on the Society of Cable Telecommunications Engineers (SCTE) Web site at www.scte.org - Don Shaver of Texas Instruments marvels at the great strides made over the past several years.

"There's been tremendous improvement in software programming productivity since the early 1980s," he says. "At that time, it might have taken 30 programmers to complete a seismic application; now, I could do the same application with about three programmers in half the time."

The tools were crude With nearly 30 years of experience in programmable digital signal processing, Shaver, who is now director for TI's communications systems laboratory, has witnessed this progression.

"Real-time embedded digital signal processors (DSPs) came into being for application in the '70s," he says. "That's when systems were widely deployed in the defense industry."

Shaver was involved in programming such a system developed by IBM early in the '70s.

"The tools were fairly crude, but there were still many advantages to using programmable DSPs to get the systems out," he adds.

Loud fans According to Shaver, in the early days, technologists required tremendous flexibility in their new systems.

"We had to continue to change the out ribbon through the life of the system," he says.

The typical system, he recalls, consisted of a control processor, which was based on a derivative of the IBM 360 processor, and had 16 kilobytes (KB) of memory. It also had a high-speed storage controller that moved data between the bulk memory and the scratch pad memories in the DSPs themselves at about 40 megabytes (MB) per second.

"It was really pushing state-of-the-art," says Shaver.

Still, he adds, "You really couldn't talk in the same room where the processor was because the fans were so loud."

Lower power for higher-volume defense A technology that came on the scene toward the late 1970s was the I2L, which rated injection logic.

"This was one-volt logic," explains Shaver. "At that time, we were starting to look at higher-volume defense applications requiring lower power."

Also among the chips developed during this period were the 16-foot arithmetic logic unit (ALU), 16-foot multiplier and micro-controller chips such as the I2L.

"The I2L was fairly popular back in the late '70s and early '80s," he explains.

It ran at about two MHz, and the power dissipation was 150 milliwatts per millions of instructions per second (mW/MIPS). This was a dramatic change from previous processors that ran, on average, at nearly 50,000 mW/MIPS.

Still, according to Shaver, the I2L never really "bore a lot of fruit."

CMOS the popular favorite By the early 1980s, the complementary metal oxide semiconductor (CMOS) had made its debut and soon became the popular favorite.

"It had a lot of nice features," says Shaver. "If you clocked it slower, you could effectively lower the power; with the I2L, the power was constant no matter what the clock rate."

Also, the one-volt I2L required special supplies for powering, whereas the CMOS, at five volts, was easier to power.

Single vs. multi-chip processors The first single chip DSP was the TMS 32010, which ran at about 5,000 MHz and had a power ratio of about 250 mW/MIPS. By 1987, Shaver and his team had developed a backward compatible version of that - the TMS 320C25 DSP, one of the most popular DSPs ever to come along. It ran at about 20 MHz, and the power ratio was 20 mW/MIPS.

"Over just a five-year period, we significantly improved our mW/MIPS power ratio," he says.

Regardless, the multi-chip/bit slice processors continued to do work for both seismic and defense applications throughout the 1980s. Eventually, they were too difficult to program and died out.

"The software wasn't reusable," explains Shaver, "and we had to develop new tools for every application."

Shaver and others started to use the single-chip DSP, which yielded better development tools as it evolved.

The high-performance VLIW The 1990s saw the use of newer VLIW (very long instruction word) signal processors.

These sophisticated processors exploited parallelism to increase the number of MIPS. In stark contrast to previous models, these new processors ran at an astounding 0.9 mW/MIPS, including all the memory in the peripherals on the chip.

Today, says Shaver, DSPs are becoming much more generic in capability and continually easier to program.

"They do much more than just signal processing functions," he says, "and can be particularly useful for cable modems, single-line digital subscriber lines (SDSL) and other communication applications."

To learn more about programmable chips, visit the SCTE "Online Lecture Series" archives at www.scte.org. The site contains other lectures, also focused on advancing technologies, from leading industry experts.