The day C6 was born:
December 28, 1998, Issue: 1041
Section: Times People 98 -- DSP Chips: Ray Simar
------------------------------------------------------------------------
Nice guy finds a DSP firestorm
Stephan Ohr
As Ray Simar tells it, he was in his hotel room, on a business trip, when it happened-a revelation that was to change the DSP industry forever. Simar is a Texas Instruments Fellow, in charge of advanced DSP architectures. "I can't tell you what hotel or what city I was in," he said. "A lot of ideas come to you when you're not at your desk.
"I had been listening to members of our team-architects, computer people, circuit designers. The compiler people were telling me their concerns (they wanted more orthogonality). . . . It's really helpful talking to different people and hearing different ideas [when you're doing a project]. Sometimes you gotta let those concerns bounce around in your head for a while. I was taking a break in a hotel when it came to me-we have to do VLIW stuff!"
Simar had remembered very long-instruction-word (VLIW) architectures as a mainframe computer technology-"really over the top." But in the quiet of his hotel room, it seemed the solution to a number of problems TI was hoping to solve with its next-generation DSP chip. The part needed to be high-performance and cost-effective. "We needed to ratchet up the compiler effort," Simar recalls, "to ease the programming requirement."
But even after his revelation, months of intensive work remained. "We had to ask what worked and what didn't work," he said. "You can't just take an array processor and put it into a bit-slice processor chip-it doesn't just move over. It had to sell for $25. We had to reshape it to see that it met all of our goals."
As principal architect of the TMS320 C6000 series, known as the C6x, Simar is credited with the design of the DSP industry's most powerful-and most controversial-DSP machine. This is the one that gave renewed life to single-chip embodiments of the very long instruction word. VLIW advocates believe increasing the clock speed is not the only way to increase the work performed (instructions executed) by a digital signal processor. It is also possible to double, triple and quadruple the number of instructions cranked through at each cycle. The C6x has eight parallel execution units, driven by a parallelizing instruction compiler. With a 200-MHz clock, the first implementation of the C6x (the C6201) was capable of cranking through 1,600 Mips.
Simar's implementation gave new life to a concept that seemed to suffer under media-processor developers like Chromatic Research and Philips' TriMedia, which are trying to harness VLIW for video and graphics (without 100 percent success). His design set off a firestorm of controversy. Critics of the approach said the C6x was a power hog; that it used too much memory. They said that its parallelizing compiler was not half as efficient as TI claimed; that 1,600 Mips was a peak figure that could never be obtained in practice; that Mips wasn't even a good way to evaluate DSP performance. Supporters of the C6x call it the first DSP that ensures high performance for programs written in C. It is the first to significantly improve time-to-market by eliminating the need for assembly-language programming by PhD mathematicians. Advocates say it's the first device to support highly parallel advanced telephony applications like T1/E1 line formatting, ADSL and digital cellular basestation control. TI claims that the C6x has shown the most rapid acceptance among users of any previously introduced TI processor.
Regardless of whether you're among the hundreds said to have endorsed TI's C6x in new designs, whether you've copied long- or variable-instruction-word thought processes into your own designs (as did the StarCore 440, Philips' R.E.A.L. and Siemens' Carmel), or whether you remain a vehement critic of the approach (as do superscalar advocates Analog Devices and ZSP Corp.), there is no argument that VLIW is the most important concept to have rocked the DSP world over the past two years. Depending on your view of the controversy, Simar, as VLIW's architect, is either the DSP industry's Man of the Year or the guy with the bull's-eye on his chest.
Simar did not realize the VLIW architecture would create so much controversy. He said he had to develop a thick skin. And he comforts himself with a long-term perspective. "Five years from now, it will all seem 'obvious'," he believes.
Simar says he took solace from the members of his team at TI-and from the conferences the DSP and microprocessor industries use to share ideas. "We had a good team of folks," he said of the crew that built the C6x. "Personally, it's been really rewarding to go through that project-and the aftermath-with them. I think our team's more tightly knit than before.
"There's ordinarily a lot of competition in the industry," he said, "but when you get people together in the same room at a conference, they're very nice and often very supportive." At a panel at DSP World, for example, Jim Boddie credited TI's accomplishments in DSP. Boddie, of competitor Lucent Technologies, is the director of the joint Lucent-Motorola StarCore Design Center in Atlanta. "That was a trip," said Simar about Boddie's compliments. "He's a real gentleman."
'Sense of community'
Simar likes to attend conferences like DSP World as a means of sharing information with friends and peers in the engineering community. "People won't spill their guts to you, but there's a real sense of community," he said. "We've been around the world together. People share ideas, try to figure out what's the next big challenge-it's always a lot of fun to be at those things."
Simar said he enjoyed the Programming Language Design and Implementation Conference, where some 300 of what he called "compiler gear heads" got together to discuss the challenges and techniques for making compilers smarter. "We want to make it easier to embed things in the compiler, like 'power intelligence,' which gives you the option of optimizing a design for power rather than just performance." Engineers there were hungry for new information, he said. "They were getting over the feeling that everything that could be done had been done."
The Microprocessor Forum also proved to be an eye-opener, Simar said. "It looks to me like there's more stuff in embedded processing. It's not just X86." He said it was refreshing to see compiler designers take a real interest in the embedded processor challenge-"that's a change."
Naturally, as a famous DSP architect, Ray Simar is very much in demand as a speaker. Recently, he delivered the keynote address-"DSP for Tomorrow"-at a Dallas conference called Micro 31. He said he offered "an assessment of some of the challenges that need to be met-just to start them thinking."
Simar sees cable modems as one design challenge: "If you look at the guts of the thing, it turns out you need a ton of Mips-tens of thousands of them. What'll be interesting is that it needs to consume only a handful of watts. So the big challenge is taking the power down." But the ability to program in C, and getting to market quickly ("you know, stuff like that"), will always remain a challenge.
Like other engineers with time-to-market pressures, Simar does not see a clear boundary between his home and professional lives. "Sometimes, you've got to get away," he conceded. But concentrated time for his hobby-woodworking-has eluded him of late. "If I could build something for the home-a wall unit, or a deck for the hot tub-I could use a little bit of my skill to create something the whole family would enjoy."
Simar may indeed have that kind of influence on other engineers. His team's experiences with the C6x drew its members closer together. And Simar said he sees threads of his work on very long-instruction-word architectures popping up in other DSP designs. "I feel kind of good about that," he said modestly. "That's all positive stuff. In this industry, things move quickly-and this'll get copied. I'm glad that people find something of value in what I've done."
Forwarded by a friend
norden |
