The US Department of Energy’s National Renewable Energy Laboratory (NREL) has ordered a $10 million HP supercomputer equipped with the latest Intel Xeon CPUs and Xeon Phi coprocessors. When completed in 2013, the system will deliver one petaflop of performance and will take up residence in one of the most energy-efficient datacenters in the world.

The supercomputer will be built in phases with the initial rack of servers scheduled for deployment this November. The first phase will use HP’s ProLiant SL230s and SL250s servers. These will be equipped with the current “Sandy Bridge” Xeons, specifically the new E5-2670 CPUs (8-core 2.6 GHz, 115W). At least some of the SL250s boxes will also host the upcoming “Knights Corner” coprocessor, the first commercial chip in Intel’s new manycore Xeon Phi line. These are due out before the end of the 2012.

The second phase of the HP system will incorporate next year’s “Ivy Bridge” Xeons, built on Intel’s latest 22nm technology. When completed in the summer of 2013, the HP cluster will house about 600 Xeon Phi coprocessors and 3,200 Xeons. Although that’s not a particularly high ratio of accelerators to CPUs, it’s likely that the vector-heavy Xeon Phi silicon will deliver more than half of the total flops for the machine.

While petascale computers are still relatively rare, the more important theme here is energy efficiency. Both the computer and the NREL datacenter (known as the Energy Systems Integration Facility) were designed to minimize power usage. At a cost of $135 million, the new facility, which includes labs and office space, is built to take advantage of the latest warm-water-cooled servers. A big advantage of this technology is that it requires only evaporative coolers for the plumbing. No chillers or mechanical cooling apparatus are needed, reducing power requirements significantly.

According to Steve Hammond, NREL’s Computational Science director, that will make it the most energy-efficient HPC facility in the world when it’s commissioned at the end of September. “We’ve taken a chips-to-bricks approach to datacenters,” Hammond told HPCwire. “We’re managing both the bytes and the BTUs.”

Since a megawatt of electricity costs around a million dollars a year in the US — and even more in Japan and most of Europe — significant savings can be achieved if these facilities can pare down their power consumption. The NREL facility was designed with that goal in mind and is targeting a PUE (power usage effectiveness) of 1.06. So for every unit of power delivered to the computing equipment, only another 0.06 more units will be needed for cooling, power supply losses, and other overhead.

For a large datacenter, that’s nearly unprecedented. According to an EPA study in 2009, the average datacenter was running at a PUE of 1.91. In these facilities, ever watt of power consumed for computing required nearly an additional watt for cooling or was otherwise wasted on transmission losses. As a result, more and more centers are turning to warm-water cooling.

Warm is the keyword here. Intake water for the computing equipment is around room temperature — 75F or thereabouts. Water exiting the servers is approximately 95F and, at NREL, will be recycled to heat the facility. Hammond says that in the future they plan to export the server-warmed water to other buildings on the rest of the campus.

NREL will not only save a nice chunk of change as a result of the energy savings, but the project will also be a showcase for PUE-minimizing design. The power-saving theme also dovetails nicely with the DOE lab’s mission, namely to support research in renewable energy and new energy sources. The HP super will be used to run computer simulations for developing clean energy, advanced solar photovoltaics, wind energy systems, electric vehicles, and renewable fuels.

With regard to the power profile of the petaflop system, HP plans to deliver a full peak petaflop with just a single megawatt. Although that’s not in the same league as an IBM Blue Gene/Q (which delivers well over 2 peak petaflops per MW), its on par with the most efficient GPU-accelerated supercomputers deployed today.

That’s due in no small part to the Xeon Phi coprocessor, which will contribute significantly to the system’s overall energy efficiency. Although Intel has not made public the wattage and performance on the initial Knights Corner chips, they are expected to be competitive with the latest GPUs, in other words, well over a teraflop of double precision number-crunching in under 300 watts.

To uphold the PUE rating of the NREL facility, the HP servers will be primarily warm-water cooled. Not only will that save energy, but it’s also the most practical approach for a petaflop supercomputer that, in this case, is being squeezed into just 1,000 square feet of floor space. The datacenter itself is 10 times that size, but this will give NREL plenty of room for disk and tape storage, not to mention additional HPC systems down the road.

In fact, Hammond says they plan to use the new datacenter for the next two decades, which should take them well into the exascale era. Since the facility can only tap 10MW, NREL will have to wait until those exa-systems fit into that power envelope. The first exaflop supercomputers are expected to draw at least 20MW when they first appear toward the end of this decade.

For now though, the DOE adds yet another petascale supercomputer to its growing roster of elite machines. At $10 million per petaflop, even smaller labs, like NREL, can now tap into computing power that was unheard of just five years ago. In 2008, the first petaflop supercomputer, Roadrunner, cost more than 10 times as much as this HP machine and took up six times as much datacenter real estate. In a few more years, these petaflop systems should be cheap enough and compact enought to be acquired by commercial users. And if these energy-saving technologies continue to be refined, such systems should be relatively inexpensive to run as well.