Intel Corp. announced several changes Monday to its road map for server processors, delaying its first dual-core Itanium 2 processor and replacing a future multicore Xeon processor with a new design that eliminates the performance penalty of shared connections to a chipset.
Montecito, the dual-core version of the Itanium 2 processor, will not be available in large volumes until the middle of next year, instead of the early part of next year as originally planned, said Scott McLaughlin, an Intel spokesman.
While preliminary shipments of the processor are already under way, Intel decided to make a few changes to the chip in order to reach the company’s standard for “production level quality,” McLaughlin said, declining to specify the nature of the changes.
But Montecito will no longer ship with Foxton, a sophisticated power-management technology, and the speed of its front-side bus connection to memory will run at 533MHz instead of the 667MHz speed originally scheduled for the design, he said.
Intel also has killed Whitefield, a multicore Xeon processor for servers with four or more processors, McLaughlin said. It is being replaced by a new processor called Tigerton that will appear in 2007, the same time-frame in which Whitefield was expected to arrive.
Tigerton processors will use a high-speed interconnect technology that will allow each processor to connect directly to the server’s chipset, McLaughlin said. Current Xeon processors in multiprocessor servers must share a front-side bus connection to the chipset in order to access data from system memory or I/O, a bottleneck that industry analysts have blamed for the current performance gap between Intel’s server chips and Advanced Micro Devices Inc.’s Opteron processors.
Intel’s next-generation architecture, announced by President and Chief Executive Officer Paul Otellini in August, will be used as the blueprint for Tigerton. This architecture is based on low-power design principles used to build Intel’s Pentium M processor for notebooks.
Whitefield had been expected to help tip the performance battle back toward Intel in 2007, but Tigerton should be even more powerful, McLaughlin said. AMD has enjoyed favorable reviews from industry analysts, and even companies such as Hewlett-Packard Co., for the performance of its Opteron server processors as compared to Intel’s Xeon chips.
Intel is not specifying exactly how the Tigerton processors will connect to the server’s chipset, such as whether they will use integrated memory or I/O controllers or a next-generation interconnect technology that Intel has vaguely discussed at previous Intel Developer Forums.
The Caneland platform, or the combination of Tigerton and its related chipset, is not the design that will bring socket compatibility to Intel’s Xeon and Itanium processors, McLaughlin said.
Intel wants to make a chipset that can accommodate either a Xeon processor or an Itanium processor, which will help reduce product development costs for both Intel and its partners. That compatibility is slated to arrive along with Tukwila, a multicore Itanium 2 processor now scheduled to arrive in 2008 as a result of the Montecito delay, he said. Intel is still evaluating when its Xeon processors will be designed for that compatibility.
Intel had been expected to introduce an integrated memory controller design at the same time it engineered the compatibility between Itanium and Xeon, said Nathan Brookwood, principal analyst with Insight 64 in Saratoga, California. Intel had never publicly confirmed those plans, but the company has spoken in general terms about the need for integrated memory controllers at some point in the future.
An integrated memory controller means the logic responsible for coordinating the exchange of information between the processor and memory is built right onto the processor, allowing it to run at the same speed as the processor and improving overall system performance. Since Intel would have had to make significant changes to its processor and chipset designs in order to make the Xeon and Itanium processors fit into the same chipset, it was considered a logical time to introduce an integrated memory controller, he said.
AMD’s Opteron uses both point-to-point interconnects like Tigerton’s and an integrated memory controller. But Intel has been reluctant to embrace integrated memory controllers, said Dean McCarron, principal analyst with Mercury Research Inc. in Cave Creek, Arizona. Leaving the memory controller on the chipset allows Intel to more easily accommodate changes in memory standards, because it’s much easier to change a chipset design than a processor design, he said.
One reason for moving out the common architecture target date, and therefore the potential integrated memory controller, could be that Intel plans to increase the size of cache memory on its future processors, McCarron said. Cache memory stores frequently used data right next to the CPU (central processing unit) where it can be accessed much more quickly than data stored in the main memory chips.
The combination of larger cache memory and the direct connections between Tigerton processors and chipsets could provide a significant performance boost for Intel-based servers in 2007, McCarron said. It’s very early to know for sure, with even sample chips still far away, he said.
However, Tigerton will certainly be more efficient than Whitefield because of the new interconnect design, Brookwood said. “What happens on the chip matters a lot, obviously, but ultimately the performance of these chips is constrained by how fast you can feed them data,” he said.