When the dust settles after its acquisition of the System x business from IBM, Lenovo Group will probably end up with the second largest HPC systems business in the world, behind Hewlett-Packard and eager to close that gap just as it has in the PC market. Lenovo came out swinging at the SC14 supercomputing conference in New Orleans this week with a new water-cooled NextScale modular system sporting over-clocked “Haswell” Xeon E5-2560 v3 processors and positioning this machine to accept Intel’s “Knights Landing” Xeon Phi processors when they ship in the second half of next year.
Lenovo is also starting up an HPC innovation center in Stuttgart, Germany, to leverage the expertise in key supercomputing centers in Europe to drive its roadmap going forward.
Before Lenovo acquired the System x division, IBM was working on a water-cooled variant of the NextScale modular system that would allow for dense packaging of compute and memory and also allow for warm water cooling of the system. Energy efficiency is a particular need for supercomputing centers, and in Europe, where electricity is more expensive and so is real estate (speaking very generally), wasting energy and space is less of an option. The NextScale system with water-cooled technology (it is literally called WCT by Lenovo) is the culmination of that engineering effort.
The NextScale systems were launched in September 2013 and pack a dozen single-width server nodes in a 6U enclosure. Unlike blade servers, the NextScale machines do not include all of the bells and whistles and redundant features that enterprise customers expect. With HPC and hyperscale customers alike, customers do not want any unnecessary feature in the box, which costs money and adds complexity to the system and therefore increases the odds that the system will break. The water-cooled NextScale machine is based on the single-wide nodes, and as Scott Tease, executive director of high performance computing at Lenovo explains to HPCwire, they include a dripless quick-connect system that runs warm water through water blocks to cool both the Xeon processors and their main memory.
IBM has been putting water blocks on processors and main memory for years, so this is nothing new. But the way that the memory is cooled a bit different from the way IBM had done it. The main memory modules in the system are spaced apart a little and a water pipe is run between them. A heat-conducting piece of metal is pushed between a pair of memory modules to suck the heat out and transfer to the water pipe. A plastic cover is placed over the memory stick pair to ensure that no air gets in to absorb the heat. While the picture above shows a bezel with holes in it in the front of the server, in the production NextScale WCT systems this front is blocked off to prevent air from coming into the system; there is a set of baffles in the back of the server that do the same as they wrap around the water intake and output pipes.
By removing the fans and water cooling the processors and main memory in the two-socket NextScale node, Lenovo has a bit of extra headroom to boost the performance of the system. And so Lenovo has done the cool thing all the server kiddies are doing these days, and that is go to Intel to get a custom variant of the new Haswell Xeon E5 processor, the family of which was announced in early September. Lenovo has a version of the chip called the E5-2698A v3, which is a 16-core chip that runs at 2.8 GHz and that can have a sustained Turbo Boost speed of 3.2 GHz. This overclocking is made possible, Tease explains to HPCwire, because the efficiency of the water-cooling leaves about 6 percent thermal headroom over air-cooled versions of the chip that run at a lower 2.3 GHz clock speed normally. After the removal of fans (which consume power) and water-blocking on the memory, the overall reduction in system power is with the NextScale WCT is even higher than 6 percent, and with the water-blocking most datacenters using this system in most regions of the world can be cooled without the need of water chillers. Demonstrating the exponential relationship between clock speed and heat, cutting the thermals on the chip by 6 percent allows a 500 MHz increase in core clock speed, with another 400 MHz on top of that with Turbo Boost pushed up and held steady.
The NextScale WCT node has eight memory slots per socket, which is more than enough for most HPC applications. With Turbo Boost jacked up on a sustained basis, the NextScale WCT node with two of the Xeon E5-2698A v3 processors has a peak theoretical performance of 1.083 teraflops doing double-precision floating point math, which is about the same performance as Intel’s current “Knights Corner” Xeon Phi coprocessor card. That works out to about 2.45 gigaflops per watt, says Tease, which is on-par with the energy efficiency of many hybrid CPU-accelerator systems. That works out to 181.9 teraflops in a standard 42U rack.
Add up the savings that come from outside air cooling, and the NextScale WCT setup can cut the energy bill by as much as 40 percent, says Tease. That means either lowering the electric bill or being able to allocate that much more electricity to a more powerful system in the same energy budget. The NextScale WCT accepts water coming in at 45 degrees Celsius, which is about 113 degrees Fahrenheit. Waste heat comes out of the rack through water that is around 55 degrees, which is about 131 degrees and plenty useful.
But Lenovo is not stopping there. The NextScale WCT will be equipped with Intel’s forthcoming “Knights Landing” Xeon Phi coprocessor, which will have a peak theoretical performance in excess of 3 teraflops per chip. Intel is making the Knights Landing chip available in its own socket, and Lenovo plans to put two of them in the NextScale WCT nodes, that will deliver at least 500 teraflops per rack. That means Lenovo could cram a petaflops of performance in two server racks. Tease was quick to point out that the “Roadrunner” system built by IBM for Los Alamos National Laboratory in New Mexico for around $100 million was the first system to break the petaflops barrier and needed 296 racks and 3 megawatts of power to do so.
We have come a long way in terms of compressing this level of performance into smaller spaces and power envelopes, and next year will be a big year for hybrid computing to push the barriers even further.
Lenovo will be installing a NextScale WCT system at the Leibniz-Rechenzentrum in Munich, Germany, and it will be Phase 2 of the SuperMUC system.
Separately, Lenovo will be working with LRZ and a number of other key players to start an innovation center in Stuttgart, Germany. It will be installing an air-cooled NextScale cluster with 5,000 cores to start. The system will use 100 Gb/sec EDR InfiniBand networks from Mellanox Technologies to hook the nodes together, and it will eventually be expanded with the water-cooling options next year. This air-cooled system will be the focal point of HPC research in Europe for Lenovo, and the system will be up and running by January with the water-cooled upgrade and the addition of Xeon Phi coprocessors coming later in 2015.