With cost considered the last battle-ground in the programmable logic market, the two largest PLD suppliers have begun a volley of verbal attacks regarding perceived manufacturing weaknesses that each says could undermine the other's ability to compete on price.
The skirmish broke out as Xilinx Inc. sent word that it will produce its low-cost Spartan-3 FPGAs using advanced 90nm process technology. This, combined with 300mm wafers, will allow Xilinx to deliver 1 million system gates, or roughly 17,000 logic cells, for less than $20 by the time Spartan-3 devices enter production at the end of 2004, the company claimed.
The volume that Spartan-3 is expected to drive will allow Xilinx to squeeze out cost and reach price points competitive with ASICs, ASSPs, and microcontrollers, according to Sandeep Vij, vice president of worldwide marketing at Xilinx, San Jose.
"People have questioned why we would introduce a low-cost family on 90nm technology, and the simple answer is that the significant die-size reduction you get allows for significant cost reduction per die," Vij said.
"Five years ago, Xilinx and Altera were about the same size in revenue," he said. "Xilinx took the strategy of getting to the newest process technology first. Altera wasn't able to do it, and now they are 60% of Xilinx's size. Which strategy do you think worked?"
Pointing to the high cost of 90-nm production, which is not considered a mainstream technology, Altera Corp. characterized the advanced process as contradictory to Xilinx's low-cost strategy. Rather, Altera said it plans to continue producing competing Cyclone parts on a 0.13-micron process for the next two to three years. Altera expects to introduce its high-end Stratix parts on 90nm by the first half of 2004. A move to 300mm wafers is set for later this year.
While analysts said a sound business case could be made for either approach, the longtime rivals pointed to manufacturing glitches that could weaken their opponent's offense.
"It looks like Xilinx is gambling on an immature 90nm process because they're having great difficulty making 0.13-micron work," said Erik Cleage, senior vice president of sales and marketing at Altera, San Jose. Cleage was referring to recent reports that Xilinx switched its 0.13-micron process to a lower-grade dielectric material after the SiLK low-k used by IBM Microelectronics produced low Virtex-II Pro yields.
Virtex-II Pro is in full production on 0.13 micron without low-k dielectrics, according to Babak Hedayati, senior director of product marketing at Xilinx.
Hedayati, in turn, argued that Altera is stuck with 0.13 micron because its sole foundry, Taiwan Semiconductor Manufacturing Co. Ltd., has reportedly fallen behind in the race to 90nm.
According to Cleage, Altera will wait to switch until 90nm die yields improve. "If you get only half as many die, the die cost has actually doubled," he said. "Xilinx's price points, not being consistent with the cost structure, will not be supportable."
How unit pricing will be affected remains to be seen, and customers are watching closely.
"Price is our No. 1 concern," said Richard Beckert, hardware design lead in the automotive business unit of Microsoft Corp., Redmond, Wash. Micro-soft is using Spartan-IIE devices in an existing telematics design that ships in excess of 500,000 units a year.
"Assuming I could do the same design with either Altera or Xilinx, I would go with the one that offered the lowest cost," said Beckert, adding that he hopes to employ Spartan-3's embedded DSP multipliers to make his existing FPGA design more efficient.
However, if Altera later introduced a more competitive part, the cost savings would be well worth the two to four weeks of design engineering time it would take to switch architectures, Beckert said.
Technology aside, Xilinx's early move to 90nm should provide a better return on investment, said Cary Snyder, an analyst at Forward Concepts Co. in San Jose.
However, that doesn't guarantee huge market share gains for Xilinx, according to Richard Wawrzyniak, an analyst at Semico Research Corp. in Irvine, Calif.
"It's really the early days for manufacturing at 90nm. There's quite a bit of a learning curve and experience that will get built up," he said. "Once Altera is ready to move to 90nm, they will be able to take advantage of it almost immediately. Xilinx will have to work through those issues upfront."
Xilinx declined to say when it plans to shift its higher-end Virtex-class devices to 90nm. The company is manufacturing 90nm-based Spartan-3 chips at IBM Microelectronics in East Fishkill, N.Y., and United Microelectronics Corp. in Hsinchu, Taiwan. Working samples have been produced at both fabs, Vij said.