Intel, AMD Raise Server Stakes

Next-generation chips to ratchet up system performance in corporate data centers.

Chip makers intel and advanced Micro Devices are preparing to roll out a new generation of chips that will deliver significant performance gains for the crucial servers that power enterprise networks.

Intel is expected to launch its upgraded "Bensley" server platform next month at its Intel Developer Forum in San Francisco. Meanwhile, AMD offered a snapshot of a forthcoming Opteron chip, due later this year, at last weeks ISSCC (International Solid-State Circuits Conference), also in San Francisco.

The new chips, which will all come to market in machines near the midyear mark, are designed to raise the performance bar for servers that are the backbone of corporate data centers. The chips also serve to continue competition between the two chip makers. The x86 server arena has seen intense competition of late as machines containing AMDs Opteron chip have gained acceptance among businesses, and Intel has moved to counter by upgrading its server processors and rolling out new platforms to support them.

Intel will get the next round started by introducing Bensley, which promises a higher-performing processor and greater memory bandwidth. A successor to the "Lindenhurst" platform, Bensley will pair Intels new 65-nanometer, dual-core "Dempsey" chip with its "Blackford" chip set.

Blackford doubles up on buses and memory channels. It includes twin buses for feeding data into its two processors and four channels for buffered memory modules. The buffered modules also work to lighten the load on the memory bus, thus aiding performance.

Intel will also include its virtualization, I/O acceleration and management technologies with the platform. The company is expected to introduce the platform at its developer forum, where it will state the silicon is ready for prime time.

Servers from brand-name manufacturers that are expected to adopt Bensley—companies such as Hewlett-Packard—may not arrive for another 60 to 90 days, however. One person familiar with the matter said to expect them in May.

An Intel spokesperson said the company is on track to ship Bensley in the first quarter. However, he declined to comment on the companys plans beyond that.

Intel likely will also discuss "Woodcrest," an all-new follow-on to Dempsey, at the forum. Woodcrest chips, which will consume considerably less power—about 80 watts, significantly lower than that of existing Xeons—will plug into Bensley platform servers and come out in the second half of the year, Intel officials have said.

AMD, working to maintain its momentum, will roll out a new line of Opterons at midyear. Company officials discussed one chip in the new line at the ISSCC, a dual-core Opteron that will run at 2.6GHz and offer double the memory bandwidth by incorporating a DDR2 (double data rate 2) memory controller. The new controller, and the subsequent move to speedier memory, will allow AMD to bump up server performance without necessarily increasing its chips clock speed or its power consumption, officials said in the presentation.

The new DDR2-equipped Opterons are expected to appear in different configurations—AMD offers different Opterons for single-processor, dual-processor and multiprocessor servers, for example—and to use a new socket, in addition to including AMD Virtualization Technology, formerly known as Pacifica. The 2.6GHz Opteron discussed at the show will consume 95 watts of power, AMD officials said.

"Its safe to say both companies are trying to make revisions to their product offerings that keep them competitive," said Dean McCarron, an analyst at Mercury Research, in Cave Creek, Ariz. "The market obviously has changed from a few years ago. Now that we have AMD present as a stronger player, its going to be very interesting."

By taking a different approach and choosing to build its memory controller into its chips, AMD bought itself an edge in performance, McCarron said.

AMD has seen its share of x86 server chips grow over the past several quarters. During the fourth quarter of last year, its server share increased to 16.4 percent from 12.7 percent in the third quarter, according to Mercury Research.

Although it may eventually use FB-DIMMs (fully buffered dual in-line memory modules), AMD is sticking with standard DDR2 for now. "They are taking fairly different approaches," McCarron said. "The market will tell us which one is right."

Meanwhile, Intels Dempsey chip may have a short stay as server makers adopting the Bensley platform may move rapidly to Woodcrest when that chip comes out later this year. Server makers have often done this in the past, McCarron said.

Intel will also beef up its four-processor-and-higher systems. For that market, it plans to expand its Xeon MP line with "Tulsa," a dual-core chip that packs more than 18MB of cache, including 16MB of Level 3 cache and 1MB of Level 2 cache for each processor core. Intel disclosed at the ISSCC that Tulsa, due in the second half of the year, will pack more than 1.3 billion transistors and measure 435 square millimeters.

Senior Editor Jeffrey Burt contributed to this story.

Seeing Xeon

Xeon Tulsa processor features:

  • Dual core, four threads
  • 1MB unified L2 core
  • 16MB unified L3 cache
  • 435-square-mm die size
  • 1.3 billion transistors
  • 3.4GHz, 150 watts thermal design power
  • Compatible with existing fonts