Supercomputers are super expensive, but Hewlett Packard Enterprise has announced plans to make supercomputing accessible for more enterprises by offering scaled down, more affordable versions of its Cray supercomputers.
The new portfolio includes HPE Cray EX and HPE Cray XD supercomputers, which are based on the Frontier exascale supercomputer at Oak Ridge National Labs. These servers come with the full array of hardware, including compute, accelerated compute, interconnect, storage, software, and flexible power and cooling options.
The first system is the HPE Cray EX2500, which has the same architecture as the HPE Cray EX4000 supercomputer used in building Frontier, but it’s 24% smaller to fit inside an enterprise data center. The remarkable thing is that the reduction came in the cooling system and not by using less hardware, said Gerald Kleyn, vice president of HPC/MCS/Edge Hardware Systems and silicon engineering at HPE. The EX2500 uses the same switches and blades as the EX4000 used in Frontier.
“HPC customers have a particular budget that they’re trying to fit into, and they’re trying to get the maximum performance inside of that budget. And so this lets us get into spaces with budgetary constraints that we couldn’t get into with the supercomputers before,” Kleyn said.
The EX2500 is water-cooled, not air cooled, and HPE integrated the coolant-distribution unit (CDU) into the server cabinet itself, rather than having a separate cabinet like on the EX4000 model. It not only saves space but also saves money. Kleyn said the HPE Cray EX2500 supercomputer costs a fraction of the price of the HPE Cray EX4000 supercomputer.
The next system is the HPE Cray XD2000, a 2U platform that goes in a standard 19 inch data center rack and is descended from HPE’s Apollo 2000 line. This server is either air or liquid cooled, and the main reason you would use the XD2000 rather than EX2500 is if you have the infrastructure in your data center for water cooling. Since the EX2500 is only water cooled, enterprises sticking to air cooling would use the XD2000.
HPE said it will offer compute blades for the XD2000 and EX2500 with both Intel’s 4th-Gen Xeon Scalable (aka Sapphire Rapids) and AMD’s Epyc 4 Genoa CPUs.
The XD6500 is a 5U platform with up to eight Nvidia H100 Hopper GPUs, making it similar to Nvidia’s DGX servers. The XD6500 is for advanced workloads including modeling, simulation and AI. Like the XD2000, it is a rebranding of an old Apollo product.
With eight Hopper processors, it should come as no surprise that the XD6500 is targeted towards the AI market and machine learning developer environment to help data scientists build out those AI workloads.
In addition to the hardware, the supercomputers come with HPE Slingshot, a high speed Ethernet interconnect, Cray Clusterstor E1000 to provide expanded storage with intelligent tiering, and HPE Cray Programming Environment, offering a fully integrated software suite with compliers and developer tools.
Kleyn said the Cray servers are a step above the newly announced ProLiant servers thanks to the liquid cooling. The liquid cooling allows for a denser server footprint, “which means you don’t have as many optical cables running across the data center to stitch all this together. So you can couple them more tightly together and get higher performance at a lower cost point,” he said.
Cray XD and EX servers with AMD Epyc processors are available now, with Intel Xeon Scalable slowly coming to market as Intel ramps production. They are available for outright purchase and through the GreenLake consumption model.