Global Sources
EE Times-Asia
Stay in touch with EE Times Asia
EE Times-Asia > Memory/Storage

ServerWorks chief spurns first-generation PCI Express

Posted: 02 Aug 2002 ?? ?Print Version ?Bookmark and Share

Keywords:serverworks? rambus? memory interface? PC sever chip? synchronous DRAM?

Raju Vegesna stubbornly refused to board the Intel-led bandwagon for the Rambus memory interface back in 1999, and the tough stance paid off. Vegesna's company, ServerWorks Inc., netted a commanding market share in PC server chip sets when server makers also turned thumbs-down on Rambus, opting instead for SDRAM.

Now Vegesna is digging in his heels again. He says he has done the math and decided against adopting the first generation (1G) of the PCI Express serial interconnect, which Intel Corp. is now driving for adoption in 2004. While the outcome for ServerWorks is unclear, Vegesna's stance reaffirms the independent and practical character of this EE-turned-CEO whose hard work and tough-mindedness have taken him from humble beginnings in India to head of a Silicon Valley company once valued at nearly $1 billion.

"I like to hire people who have strong opinions. He does not hesitate to let me know when he disagrees with me, and I like that," said Henry T. Nicholas, CEO at Broadcom Corp., which acquired ServerWorks in January 2001 for $957 million in stock.

Yet to be seen is how top managers at Dell Computer Corp., one of ServerWorks' major customers, will feel about Vegesna's decision on PCI Express, a technology Dell has told suppliers it wants to be the first to deploy. "The current version of PCI Express at 2.5Gbps is not offering much for the server that PCI-X cannot do," said the 40-year-old Vegesna. "At some point, Express could be ready for serversmaybe in the 5Gbps or 10Gbps generation."

ServerWorks and Dell started "a frank discussion" on that issue in early July at a private industry meeting called by Dell. The results of such dialogues may set the course for how mainstream servers evolve over the next few years.

"We went through this whole serial-bus issue with Rambus," said Vegesna. "We performed the calculations and found PC-133 SDRAM was enough. It is the same exercise we are doing with PCI-X 2.0 now."

Back in 1999, many in the industry agreed with ServerWorks' math and decided that the Rambus equation did not add up. The company's chipset orders shot up while Intel, which supplies the majority of PC chipsets, fumbled with a Memory Translator Hub that would let its Rambus chips support SDRAM. "That opened up an almost two-year gap between [ServerWorks] and Intel," said market watcher Nathan Brookwood of Insight64.

The two Santa Clara, California companies are squaring off again as Intel moves back into the server chipset business in a big way this year. They are likely to arm-wrestle over the future of server design for the next couple of years as Intel pushes hard for Express while ServerWorks rolls out its own I/O road map this fall, centered on Gigabit Ethernet and its follow-ons. Defining the server's I/O pathways is strategic. "The server is positioned between the storage-, wide- and local-area networks. It's the hub where they all meet and communicate," said Broadcom's Nicholas.

Despite the competition, ServerWorks maintains a good relationship with Intel, said Vegesna. Intel uses ServerWorks' chips in several of its motherboards and has even referred chipset sales to ServerWorks in some of its presentations to OEMs and analysts.

Vegesna, who became an engineer at the prompting of his father, says he was fascinated by electronic technology despite the fact that he owned neither a TV set nor a computer at the time he earned his EE degree from India's Bangalore University in 1983. "We saw them in the labs," Vegesna said. The following year, he learned about automotive electronics while earning his master's degree at Wayne State University, outside of Detroit, where he also took a course on the Motorola 68000. "I was amazed by microprocessors. I thought they were a powerful technology that would take over the world," Vegesna said.

He decided to create a software simulation of the 68020 for his thesis and called a senior Motorola engineering manager for help. Vegesna showed his work to the company contact when it was done and had a job offer five days later. Eventually he was assigned to write the microcode for the 68030 and the spec for the '040.

Two seconds flat

One day in 1988 Vegesna's boss, Roger Ross, called him and invited him to be chief architect for a startup he was planning, a company devoted to developing 32-bit Sparc processors. "I accepted in two seconds over the phone. He used to say he took longer convincing his wife [to marry him] than me," said Vegesna, who described Ross as "a very dynamic person. I still owe my career to what he did for me."

The 1G HyperSparc from Ross Technology Inc. was "my career achievement as an engineer," said Vegesna. "It was the most fun I ever had. We had to do everything from scratch, and we blew away every other processor in performance."

When Sun Microsystems Inc. choose the CPU as the heart of its first multiprocessing computer, Vegesna worked closely with Sun cofounder Andy Bechtolshiem and engineer Steve Kleiman (now CTO of Network Appliance) on the so-called Galaxy system. "That gave me a lot of insight on multiprocessing," said Vegesna.

About that time, Intel launched its first Pentium chip and began talking about multiprocessing, as did Microsoft with its then-new Windows NT. "I developed some technology contacts in those companies and saw what they were doing. I could see a high-volume market developing for MP machines," said Vegesna.

In 1993, after subsequent HyperSparc chips failed in the market, Vegesna and two friends struck out on their own in a startup called PRQ and later renamed Reliance Computer Corp. The 12-person company, largely funded with Vegesna's own money, developed a multiprocessing chip set for the Pentium. Vegesna ran a tight ship. "One person was a master's in engineering who ran the simulations and doubled as a receptionist and courier for parts as they arrived at the airport. He's still with me today," said Vegesna.

In 1996, Compaq Computer Corp. took a chance on the startup's 2G part for the Pentium Pro, impressed with its novel I/O caching scheme and other concepts. Gaining an IBM design win was tougher. "I had to build two working platforms to show themthe 1G Pentium and the Pentium Pro that Compaq was using. Then, after Compaq shipped, IBM said it would use our 3G chipset," Vegesna said.

By mid-1997, Reliance was profitable, and Vegesna sold a majority stake of the fledgling company to Fujitsu in order top get working capital to expand. That turned out to be a lucky break when Fujitsu helped Vegesna negotiate access to its broad 10-year cross-licensing agreement with Intel in 1998.

Fujitsu was richly repaid for its help. When Reliance, by then renamed ServerWorks, was preparing for an IPO, Vegesna bought back Fujitsu's original $4 million stake in the company for $140 million, largely in compensation for the Intel cross-license ServerWorks holds through 2008.

Small staff

ServerWorks today employs about 120 people, a small head count compared with the manpower that analysts believe Intel puts into its server chip set business. Intel hired nearly 100 server chip set designers from Hewlett-Packard alone in 2001.

Vegesna says steering through the industry's former bus wars and his own HyperSparc days have taught him a few lessons he takes into the current melee: You have to win the business with every new silicon generation. Execution is vital. And "if you blink, you fail," he said.


EE Times

Article Comments - ServerWorks chief spurns first-gener...
*? You can enter [0] more charecters.
*Verify code:


Visit Asia Webinars to learn about the latest in technology and get practical design tips.

Back to Top