Intel's Xeon 5500 series processor, the follow-on to Harpertown and the chip formerly known as the Nehalem-EP was launched this week, was launched this week and computer vendors the world over collectively exhaled their announcements onto the IT press.

This week's release of the Xeon 5500 is Intel's unveiling of the server version of the Nehalem microarchitecture, the successor to Intel's Core microarchitecture. In Intel's tick-tock release schedule, this is the tock: Nehalem uses the same 45nm process as Penryn, but introduces a new microarchitecture. The next tick, codenamed Westmere, will shrink the process to 32nm.

Intel and many of the OEMs that build on Intel's platforms are saying that the 5500 is Intel's most significant chip advance in more than a decade. Intel Senior Vice President and General Manager of the Digital Enterprise Group Pat Gelsinger orchestrated a Nehalem sales pitch to a live and remote webcast audience on Monday in San Francisco, where he characterized the new chip as "the most important server launch since the Pentium Pro" -- the company's initial server processor, introduced in 1996.

There is a lot to like about this chip for high-end computing. The three integrated memory controllers give the Xeon 5500s more than triple the memory bandwidth of the 5400s, and the chips can access 18 DIMM slots, which, when fully loaded with 8 GB DIMMs, puts 144 GB of DDR3 RAM in a system. This is also the first appearance of the new QuickPath Interconnect (QPI), which replaces the frontside bus and is Intel's answer to AMD's HyperTransport. At 25.6 GB/sec peak, QPI provides a nice foundation for building SMP nodes on a multiprocessor platform. There are other advancements in threading performance, and in components related to processor, including the new Intel 82599 10 Gigabit Ethernet Controller, which has virtualization technology in it designed specifically to boost network performance in centers that rely heavily on virtualized servers.

An interesting feature of the chips that many OEMs are taking advantage of is advanced power management. On HP's new G6 lineup of ProLiant servers, for example, you can set a per processor power cap at the system level. I asked HP's Ed Turkel why this particular feature would matter in HPC. He explained that you can put these processors in an existing datacenter that you're "pretty sure" can handle the load without worrying about exceeding the design point for the facility -- you just set the cap where you need it to keep from tripping breakers, and run worry-free (of course, monitoring power usage to make sure things are behaving as expected). And the chips use 50 percent less power at idle than the 5400s, so for those odd times that you aren't using your cluster you can expect to save a few dollars.

According to Gelsinger, as the earliest shipping segment, HPC is leading Intel's deployment of the chips. And look for the 5500 to be installed in multiple petaflops-class machines in 2009, including NASA's Pleiades supercomputer and the Canadian Scinet machine -- part of the CERN Hadron Collider project. Intel's release also sparked the HPC OEMs like IBM, HP, Dell, Sun Microsystems, SGI, Cray, Appro, and Penguin Computing to announce their new systems and upgrades that take advantage of Nehalem's features.

From Turkel's perspective, HP's new ProLiant G6 portfolio is "not just a new processor, it's a major architectural shift" for the company's product offering. The G6 line puts a strong focus on the datacenter, and Turkel says that one of the objectives of the new line was to think about performance, cost, and intelligent power utilization as equally vital to creating an effective solution for the HPC space. The new ProLiant incorporates what HP is calling a "sea of sensors" to monitor each server at dozens of points. This enables, for example, careful control of fan speeds to minimize their power draw. HP is also introducing the Common Power Slot design to allow customers to choose from four power supplies that match their specific needs without the need to change anything else in the system. The G6 line also takes advantage of HP's new Flex-10 technology, providing 10 GbE support on the motherboard that can be virtualized such that each core on a two-socket system can have a dedicated 1 GbE link. This isn't necessarily a big win for high-end HPC, but it should certainly matter to the enterprise market and smaller-scale HPC systems.

Like HP, IBM has introduced Nehalems across its server product line. The chips have their HPC impact in IBM's Cluster 1350 portfolio, and the 5500 has been integrated into new versions of the HS22 blade server, the rack servers, and the iDataPlex. Bob Galush, IBM's vice president for System X High Volume Servers and Options, emphasized the benefits of the launch and the extra engineering IBM added to its systems around Nehalem, for HPC as well as the broader computing market. "With this launch IBM has significantly improved the management, power management, and performance story of its offerings."

IBM has integrated voltage regulators on the motherboard, added counter rotating fans, and incorporated 92 percent efficient power supplies, all in an attempt to improve the datacenter characteristics of its offering. IBM has also added quite a bit of new software to make sure that its systems don't use more power than is required for a specific workload, and to help users manage hardware more effectively. The Canadian Scinet machine referenced earlier uses Intel's new Xeons as part of a hybrid architecture expected to perform at 360 TFLOPS. The system will couple an iDataPlex system and POWER6 architecture to support widely varying workloads by allowing users to match the system to the characteristics of the work to be performed.

Beleaguered HPC vendor SGI also introduced Nehalem on its Altix ICE line of clustered supercomputers. Presumably we'll see these machine show up under Rackable's brand, assuming the SGI buyout goes according to plan. SGI included benchmark results for the new machines showing significant performance improvements on complete HPC applications, not just benchmark kernels. Results for Fluent, VASP, and WRF were reported with improvements ranging from 1.73x on Fluent to 2.4x on WRF for similarly configured ICE systems running the previous generation Xeon 5400 processors.

SGI also put early eval version of its Nehalem-based ICE into customer hands. Jeffrey Vetter, Future Technologies Group Leader at the Department of Energy's Oak Ridge National Laboratory, said in a statement: "Our early evaluation of Nehalem demonstrated significant performance improvements on many of our important applications. By harnessing the power of SGI Altix ICE systems driven by Intel Nehalem processors, we expect to shorten our time to results, which in turn will give our users greater flexibility to pursue solutions to complex, 'big data' problems."

Cray's high end is still dominated by chips from AMD, but it did slot the new processor into its CX1 line of deskside supercomputers, where Ian Miller, senior vice president of the productivity solutions group and marketing at Cray, sees the combination of the new chips and the CX1 form factor helping drive "drive high performance computing farther into the mainstream."

Penguin Computing has added the 5500 to its Relion 1700 and 2700 series systems and, according to CEO Charles Wuischpard, also managed to get early-access Nehalem gear into the hands of customers. Chip Watson, deputy CIO and head of High Performance Computing at Thomas Jefferson Labs, commented on the new gear in a statement released by Penguin along with its announcement. "At Thomas Jefferson Laboratories we have been running particle physics simulations on a new Penguin Relion 1702 system. Results are very encouraging as we have been able to achieve more than 2x performance increases in our processor and memory bandwidth bound codes."

Finally, HPC and enterprise manufacturer Appro, who recently announced a large deal with the DOE, is using Intel's new chip in its "HyperGreen" cluster. From the branding you might guess that the company is emphasizing the power management aspects of the cluster, and you'd be right. The solution puts 640 cores in a standard (42U) rack, and the company says HyperGreen saves 20 percent over other 1U servers. Do note, however, that each node only supports up to 48 GB of DDR3 memory, not the full 144 GB allowed by the Nehalem architecture. This could be a downer for those with large memory requirements.