Intel fights back with high-end Xeon
New "Tulsa" chip for high-end x86 servers takes on Advanced Micro Devices with onboard cache memory.
- Shankland covered the tech industry for more than 25 years and was a science writer for five years before that. He has deep expertise in microprocessors, digital photography, computer hardware and software, internet standards, web technology, and more.
The new dual-core chip, designed for four-processor systems and officially called the Xeon 7100 series, has a price ranging from $856 for a 7110 model with 4MB of high-speed cache memory and a 2.6GHz clock speed to $1,980 for a 7140 model with 16MB of cache and 3.4GHz speed.
Last week, Intel said Tulsa boosts performance about 70 percent compared with its predecessor, "Paxville," but now the chipmaker is trying to steer attention toward a comparison with AMD's Opteron. Tulsa systems are 17 percent faster than Opteron machines on business database tasks and 42 percent faster on Java server tasks, Intel said.
AMD entered the x86 server market three years ago, and the competition has been fierce since. All four major server makers sell or will sell Opteron servers, and AMD has risen to claim 26 percent of the server processor market. Intel is fighting back with its "Woodcrest" Xeon for dual-processor servers and now with Tulsa for higher-end models.
"The 7100 really brings us up back to where want to be in terms of maximum performance," said
Tulsa is the last model to use Intel's all-but-extinct NetBurst architecture, which Kilroy acknowledged is "long in the tooth." But Tulsa still is a compelling design, he argued. "The 16MB of L3 (level-three) cache really helps position it for superior performance."
Cache is king
Indeed, cache is at the heart of the AMD-Intel rivalry. AMD's Opteron includes a built-in memory controller, while Intel systems require a separate chip that takes longer to fetch data. But including a large cache means that data is more likely to be readily at hand, so the memory controller isn't needed at all, Intel argues.
Intel's 65-nanometer process means more circuitry can be squeezed into a given surface area than with the 90-nanometer process AMD still uses. And even using the same process, Intel cache elements are smaller, said Pat Gelsinger, who co-manages the Digital Enterprise Group with Kilroy.
"My cache cells are about half the size. I got cache to burn," Gelsinger said in an earlier interview.
Tulsa also is the first Xeon to include "Pellston," officially called Intel Cache Safe Technology, which shuts off cache elements if errors are detected. Such reliability features are important in higher-end servers, Kilroy said.
But cache isn't free. It takes up real estate, increases manufacturing costs and makes a chip run hotter. The top-end Tulsa with 16MB of cache draws 150 watts of power running flat out, though slower models with 4MB cache consume only 95 watts.
AMD was quick to point out the power difference compared with its 95-watt mainstream Opteron chips. "We don't require customers to choose between high performance and great power savings," said John Fruehe, worldwide market development manager for AMD's server products.
More feistiness is all but guaranteed in the x86 server market, which generated $5.9 billion in revenue for computer makers in the second quarter, according to IDC.
Server makers are happy to have two x86 server chip suppliers. Several announced new products based on Tulsa:
• Dell has begun selling Tulsa-based PowerEdge 6800 and 6850 servers with a starting price of $6,900. The 6800 is a free-standing model, and the 6850 is rack-mounted.
• Hewlett-Packard will include Tulsa in its ML570 and DL580 servers. "These should be available right around the same price points as (models with) the previous-generation chips, giving customers added performance at the same price," said John Gromala, director of server product marketing. For example, the rack-mounted DL580 with dual processors and 4GB of memory costs $8,297 and a similarly configured free-standing ML570 costs $7,477.
• IBM will bring Tulsa to its rack-mounted four-processor System x3850 in late October. Its free-standing four-processor x3800 models will ship with Tulsa in November, the same time frame as the x3950, a system with a four-processor chassis that can be linked into a single 32-processor system.
• Unisys will use the chips in its ES7000 servers beginning in the fourth quarter, at which point the company will announce prices.