Back at the International Supercomputing Conference in June, supercomputer maker Eurotech dropped some hints about its future water-cooled Aurora systems that would employ a mix of ARM processors and Nvidia Tesla GPU accelerators in a dense form. At the SC14 conference this week, these machines have now been officially launched as the Aurora Hive systems, and it turns out that the systems will also allow customers to build massively parallel machines based on Intel Xeon processors and Xeon Phi coprocessors.
The Hive systems use a modular enclosure that that is based on a cubic shape rather than a hexagonal one, but the concept of densely stacking compute elements while isolating them from each other, as a beehive does, holds true. The system crams up to 128 nodes (which are called bricks) into a single rack – 64 nodes in the front and another 64 nodes in the back, which is something you can do when you use water cooling on the components of the nodes because you do not have to worry about airflow from cold to hot aisles through each rack.
The Hive system makes use of a second generation of direct hot water cooling from the Aurora line, which Fabio Gallo, Eurotech HPC business unit managing director, tells HPCwire can cool a system with 50 degree Celsius (122 degrees Fahrenheit) inlet water temperature. The new water cooling is lighter and more compact, allowing for more compute and cooling to be crammed into the same space. The water distribution system is built right into the Aurora Hive rack, and there are dripless connectors for inlet cold (relatively speaking) and outlet hot water coming off each node. Being able to take the heat away quickly and efficiently is vital because a fully configured Hive rack draws 166 kilowatts of juice.
“You can free cool this machine nearly anywhere on earth,” says Gallo. By Eurotech’s math, customers using the Aurora Hive should be able to attain a power usage effectiveness of 1.05, which is about as good as the hyperscale datacenter operators are getting. (PUE, as this metric is abbreviated, is the ratio of the power consumed by a datacenter divided by the power consumed by the compute, storage, and network components of the datacenter. Getting as close as possible to 1 is the goal.)
The Hive nodes are 3U high, and you can put them into a rack four across and sixteen high. (Each node is 130 mm high by 105 mm deep by 325 mm deep.) Each node has a system board that includes risers for a compute module and five coprocessor modules; this system board also includes a PCI-Express 3.0 switch from PLX Technology (now part of Avago Technologies) that links the compute and coprocessor elements to each other. The PCI-Express switch also has hooks out to network adapters, in this case a two-port FDR InfiniBand adapter from Mellanox Technologies. All of the PCI-Express slots have the full bandwidth of an x16 slot, which means Nvidia Tesla GPU and Intel Xeon Phi coprocessors can find a place.
Eurotech’s first Hive system will have a CPU compute element that is based on Intel’s “Haswell” Xeon E3-1200 v3 processors. This family of chips has four cores and clock speeds that range from 3.1 GHz to 3.7 GHz in standard versions. The Intel E3-1200 v3 compute node has 32 GB of memory welded onto it for low clearance and also has a 256 GB half-height 1.8-inch solid state disk drive. You can use any E3-1200 v3 chip that has a thermal design point of 84 watts or lower.
The compute brick allows for up to four coprocessors to be fitted with cold plates for sucking the heat off their components and linked to each one of the cores over the PCI-Express switch and into the PCI-Express controllers on the E3-1200 processors. Gallo tells HPCwire that it will ship the Xeon E3-1200 plus Xeon Phi configuration in a few weeks to initial customers, and that a few months after that the combination of the Xeon E3 processor and Nvidia’s Tesla K40 coprocessor will be supported. The Xeon Phi 7120X is rated at 1.2 teraflops doing double precision floating point math, while the Tesla K40 card has a base performance of 1.43 teraflops that can rise to 1.66 teraflops with GPU Boost overclocking turned on. That works out to 614 teraflops per rack with Xeon Phis and 732 teraflops per rack with the Tesla K40s (not counting the extra performance from GPU Boost).
Back in June at ISC, Eurotech was talking up the Hive system (which did not yet have that name) by saying that it would be delivering a variant of the system that would marry a 64-bit ARM processor from Applied Micro with Tesla GPU coprocessors, and you might have gotten the impression that this would come out first. While Applied Micro is shipping its “Storm” X-Gene 1 chip now, it is readying the much-better “Shadowcat” X-Gene 2 processor, which has been sampling since August. This chip will support the RDMA over Converged Ethernet (RoCE) protocol over its integrated Ethernet network interface cards, simplifying the components that go into an ARM server node. The X-Gene 1 and 2 chips have two 10 Gb/sec Ethernet ports on the die, and these can be hooked eight into adapter ports. That, in theory, leaves more room for other peripherals in the complex. The plan is to ship the X-Gene 2 as the ARM option for the CPU side of the hybrid node, along with the Tesla K40 cards as coprocessors, sometime around the second quarter of 2015.
Incidentally, Eurotech is able to get its hands on a modified Tesla K40 card with its thermal plates modified so it fits into the super-skinny Hive module. The new Tesla K80 coprocessor card, announced this week at SC14, will be a bit tricky to add to the Aurora Hive system, explains Gallo, because this dual-GPU card has some of its power connectors across the top of the card. This does not work with the very tight tolerances in the Hive module, which are necessitated by the thermal conduction plates. With the Tesla K80 offering a base 1.87 teraflops of double precision math with a GPU Boost of up to 2.91 teraflops, you can bet some customers will want this. Gallo says that there is enough thermal capacity to pull the heat off this 300 watt part, if the connectors can be sorted. Being able to double the flops in the box is a pretty strong motivator to solve this engineering problem.
Generally speaking, the X86 processor option plus either the Xeon Phi or Tesla GPU accelerators draws about 1,500 watts per node, which works out to around 5 gigaflops per watt. The top machines on the Green500 ranking of supercomputers are in the range of 4 gigaflops per watt.
Gallo is tight lipped about what other processing components it might add to the Aurora Hive system, but obviously next year’s “Knights Landing” Xeon Phi, which will be sold as a standalone processor as well as a PCI-Express coprocessor, will slide right into this system. At 3 teraflops of double-precision floating point performance, and with the ability to put in five cards, this will be a radical increase in the math capabilities. And for dense-packed, CPU only workloads that used low-speed Ethernet, Eurotech could make Hive bricks that are just based on Xeon E3 or various ARM processors which sport their own networking on the chip. If you take out the network card, that leaves room for six CPU-only compute cards per module, or 768 processors per rack. Another option would be to add cards that have flash drives with the high-speed, low-latency NVM Express protocol linking into that PCI-Express switch. You could also swap out some of the flash drives and put in GPU cards for visualization to do visualization in the same nodes where the data is stored. Eurotech has lots of options with the Aurora Hive architecture, and that is so by design.
But initially at least, Eurotech is going after the workloads that have been accelerated. “There are markets where accelerated application have become the norm instead of an exotic thing,” says Gallo. “Geosciences, particularly reverse time migration reservoir analysis, is a good example. In general, signal processing will be interesting on this system, as well be machine learning, analytics, and some computer-aided engineering tools that have been modified for accelerators.”
The Aurora Hive comes preconfigured with the CentOS 6.X variant of Linux and support from Eurotech for this distribution, but customers can deploy other Linux operating systems on the machine as needed. Scientific Linux, Red Hat Enterprise Linux, SUSE Linux Enterprise Server, and Canonical Ubuntu Server are all supported. The Aurora software stack includes support for Intel Cluster Studio, Nvidia CUDA, MPSS, and the GCC compilers as well as the Intel MPI, Open MPI, and MVAPICH2 communication libraries.
Pricing for the Aurora Hive system was not available, and the question is what kind of premium can Eurotech charge for density and hot water cooling. The combination of the two should allow Eurotech to command a premium for its systems over plain vanilla clusters based on rack or blade servers, but it is a question as to how much. The market will decide.