Apr 23, 2002 Austin, TX
Intrinsity Inc. has spent five years quietly developing dynamic logic, a creative twist on static CMOS logic. Now, the 90-person startup is going public with two processors that build on those dynamic logic circuits to hit chip-level speeds that standard CMOS may not be able to match.
The first, FastMips, is a 2-GHz MIPS32-compliant processor, with two Rapid I/O ports for passing data quickly to other processors. The second, FastMath, adds a dedicated matrix-math processing engine, also running at 2 GHz, to the FastMips part.
Both processors are "within a few months of taping out," said Intrinsity president Paul Nixon. They are scheduled to begin sampling in the fourth quarter and go to commercial production on a standard 0.13-micron process early in the second half of 2003.
While it's hard to predict how competing embedded processors will perform at that time, Nixon said, he expects the FastMips clock rate to be at least double or even triple its competitors' in the embedded arena.
Thanks to its matrix-math engine, FastMath is optimized for adaptive signal-processing applications, such as radar and sonar detection, image processing, medical diagnostics and cellular "smart" antennas.
One potentially large market, Nixon said, is multiuser detection, or MUD � adaptive algorithms that perform interference mitigation in third-generation cellular basestations. MUD-enhanced basestations generally require faster processing of Viterbi and equalization algorithms than today's DSPs can perform, said Jim Gunn, a senior consultant at Forward Concepts (Tempe, Ariz.).
Gunn said smart antennas and MUD technology will be needed to cram more users onto the existing infrastructure and to offer data, video, music and other added 3G-type features.
Gunn expects the startup to face heavyweight competitors. "I have no doubt that Intrinsity has done a very, very nice job," Gunn said. "They have really done their homework when it comes to their matrix processing engine. That being said, there is going to be a lot of competition coming from TI, Motorola and Agere."
Part of Intrinsity's "homework" was done with engineers at Mercury Computer Systems Inc. (Chelmsford, Mass.), which develops wireless, military and medical systems. Craig Lund, chief technology officer at Mercury, called Intrinsity "a pioneer" in adding vector processing to a microprocessor.
Engineers in Mercury's wireless communications group "spent time working with the Intrinsity people, and we had a lot of interesting conversations over how to optimize their matrix engine for MUD," Lund said. "We may have had some influence there over what the Intrinsity matrix engine ended up looking like."
Mercury uses a 500-MHz PowerPC part in its current basestation system, so the 2 GHz that Intrinsity promises is enticing. But Lund cautioned that beyond raw performance, Intrinsity will need to hit certain power and cost targets. And like any startup, Intrinsity needs to prove that it can ship what it promises, Lund said.
'Pioneering stuff'
The dynamic logic approach used by Intrinsity "is pioneering, startup kind of stuff, but for any early pioneer to displace the established vendors they have to prove that it works in actual products. We'll know that when we have a chip in our hands," Lund said.
Nixon said basestation vendors want a programmable solution that runs as fast as the hardwired ASICs that now populate their boards. "DSPs don't handle matrix math very well � it takes tons of code to do matrix math on a DSP," he said. "The customers we talk to love the programmability of DSPs, but can't get the performance out of legacy DSP architectures. So they use ASICs and FPGAs, but those burn power and are very expensive. And when the algorithm changes they need a programmable part so they don't have to do a truck roll to upgrade."
Texas Instruments Inc. also is adding coprocessors and accelerators to its TMS320C6416 platform, technical marketing manager Henry Weichman said. Ten 3G basestations have been developed thus far, said Weichman, and the platform has been designed into eight of them. That 6416 chip includes a C64x DSP running at 600 MHz, a Viterbi accelerator for voice algorithms, and a newly developed "turbo" engine for the turbo convolutional code used to speed data processing in 3G basestations.
Nixon counters that Intrinsity's FastMath part will perform 596,000 (1,024-point, Radix-4) fast Fourier transforms per second, and rates the TMS320C6416 at less than 100,000 FFTs/s. Intrinsity's FastMath matrix-algebra coprocessor is based on an array of 32-bit processing elements, each with a local register file, which can handle fixed-point matrix, vector and scalar data types. Peak performance is 64 Gops (64 billion operations per second).
Weichman said TI has researchers working on new accelerators, but he stopped short of saying the Dallas company was specifically developing a matrix-algebra engine.
Fight over power
Weichman called power a critical difference between Intrinsity and the established vendors.
Intrinsity executives said that dynamic logic burns more power, partly because static CMOS is easier to turn off when a chip is idle. Intrinsity said that at 2 GHz, the FastMath part will burn 15 W and the FastMips 10 W.
Weichman scoffed at that power budget, saying that because so many chips are running in parallel, the basestation market demands parts that consume in the range of a single watt. Operating temperatures for the chips must be kept cool.
Hal Feeney, an analyst at Pathfinder Research (San Jose, Calif.), called power consumption important but said he is confident that customers will find uses for the performance Intrinsity promises to deliver. "What will happen is that after Intrinsity is shipping for awhile, customers will come to them and say, 'We could really use the performance to solve this problem.' But initially, Intrinsity may face the challenge of how far the customers will go to take advantage of a new technology. Even though the performance Intrinsity is talking about is very impressive, until people have the parts in hand, it is hard to say what will happen," Feeney said.
Tom Starnes, a DSP and microcontroller analyst at Gartner Dataquest, said that Intrinsity's foundation in dynamic logic may give the company "a kind of magic potion in the basic transistors that makes them much faster than others. Will they be able to maintain that lead, and do it in high-volume manufacturing? We'll see."
