X

AMD swings Hammer at server makers

The company won't find it easy to snare a deal, but low development costs might help it get a major server manufacturer to use its chips.

Michael Kanellos Staff Writer, CNET News.com
Michael Kanellos is editor at large at CNET News.com, where he covers hardware, research and development, start-ups and the tech industry overseas.
Michael Kanellos
5 min read
AMD increased its share of the PC processor market in part by lowering the costs for computer makers, and in 2003 it will try the same thing with servers.

The Sunnyvale, Calif.-based chip manufacturer is touting the benefits of its Hammer chips for servers, which will be marketed under the names Opteron and Athlon. One of the expected advantages comes down to a comparatively low cost of designing, and subsequently managing, servers built around the chip.

Unlike servers based on Intel processors, which require different chipsets for systems with one, two, four or eight processors, Athlon and Opteron servers would use chipsets that would be roughly identical regardless of the number of processors--dramatically cutting the amount of backroom engineering and design work.

In other words, a four-processor Opteron server wouldn't be a rarified machine with its own distinct architecture and chipsets; it would be a one-way box replicated four times.

"You won't have to have multiple chipsets and multiple designs," said Ed Ellett, vice president of marketing for AMD, who added that the company is having "good conversations" with server makers about using the chips. "The reality is that no one is going to cut over to AMD 100 percent, but the value proposition is high."

The identical chipsets will also make the chips cheap for corporations to adopt in that IT managers will have to be educated to manage only one basic computer, Ellett added.

It's a "fairly elegant" design, said Dean McCarron, principal analyst at Mercury Research.

The server market remains one of the prime growth areas for companies like Intel, Microsoft, Dell Computer and, now, AMD. Although PC-based servers--or servers that use chips based on designs originally from Intel--have been around for years, the traditional RISC/Unix server manufacturers like Sun Microsystems continue to snatch the bulk of the revenue and dominate the upper echelons of the market.

PC-based servers may have accounted for 88.4 percent of the servers sold in the first quarter, but that adds up to only 40 percent of the revenue, according to statistics compiled from IDC reports.

Two weeks from now, Intel will try to move further up the ladder with Itanium 2, a new processor that will be incorporated into servers containing four to 64 chips.

Similarly, AMD will seek to attack Intel underneath with the Hammer family. Opteron, for two- to eight-processor servers, and Hammer-based Athlons, for one- and two-chip servers and workstations, will ship to manufacturers in the first quarter of 2003 and appear commercially in the first half of that year--one year after originally planned.

The key ingredient behind Hammer computers lies largely in how the chips communicate with one another. In traditional servers, the processors funnel instructions through a common chipset, which connects to main memory, hard drives and peripheral devices. Because chipsets serve as the main artery between the processor and the rest of the world, thousands of engineering hours are consumed in developing and debugging them.

Delays on the most complex chipsets are fairly common. The differing performance requirements between server classes also mean that chipsets for one-, two-, four- and eight-processor servers are different from one another.

By contrast, traffic inside Opteron servers gets dispersed. Memory is located next to each processor, and the processors are connected to each other through HyperTransport links, a high-speed, chip-to-chip interconnection protocol invented by AMD. To obtain data, processors issue commands that travel around a high-speed ring road connecting suburbs of memory, rather than going on the trek across the bridge to a massive memory bank.

The processors must still share a common gateway for communicating with input-output devices, but the traffic across the common channels is greatly reduced because of the HyperTransport web connecting the processors and memory. A conceptually similar interconnect architecture is used inside IBM's Regatta server.

In addition, the memory controller, usually a separate chip, is integrated into the chips of the Hammer family, speeding data transmission even more.

Ironing out the kinks
Like Intel, AMD will work with server makers, software developers and potential customers to smooth out kinks in the adoption curve as part of its "Rolling Thunder" campaign to build support for Hammer. The company will create and freely license reference designs for building two-processor Athlon servers and will work with the major Taiwanese manufacturers to deliver boards to market.

Regional server manufacturers believe the chips will be a hit with customers.

"A lot of people are anxious for the Hammer technology," said Roland Baker, president of Net Express, which makes servers and workstations for universities, chip designers and the number-crunching departments in corporations.

A number of the company's customers have recently switched from buying AMD-based servers to Intel-based servers because Intel has dropped prices and increased speeds on its Xeon chips. However, the pendulum will likely swing back if Hammer lives up to the promises.

"People will have no hesitation about going back to AMD," Baker said. "The price is less, and they want to support the underdog. This is not us promoting it. That is what the customers are telling us."

AMD also will collaborate with the major manufacturers to work on their own designs. So far, none have publicly announced that they will produce an AMD-based server for the North American market. Some last year said they had little interest in it. But Ellett said that all of them are testing the chips in their labs.

"We have all of the top-tier server manufacturers interested in our products," Ellett said.

Landing those deals, though, will be difficult. Most of the major server makers have already invested millions of dollars in creating servers based on their own chips, as well as servers that use Intel's Xeon and Itanium chips. Some European and Japanese manufacturers have used AMD chips for servers but in limited capacities.

"The only real issue in my mind is whether AMD will be able to convince HP or IBM or someone else of that ilk to sell them," said Nathan Brookwood, an analyst at Insight 64. "They've got to score some success in the enterprise space or at least the small-business space."

The company has won backing from Microsoft for the Hammer chips.

Despite the challenges, Brookwood complimented the company on its homework. "They've put a lot of effort into the platform aspect of Hammer as well as the instruction-set architecture," he said.

Major manufacturers, moreover, will likely use the Hammer version of Athlon first, for one- and two-processor servers, and not Opteron, for four- and eight-ways.

Still, it will take only one of the major manufacturers to find success in switching to begin generating more widespread interest, McCarron said. Besides, the regional specialists and system integrators, along with the "white box" manufacturers in the developing markets, are growing at the expense of the major brands.

"The white-box guys aren't doing any independent R&D," McCarron said.