The era of the one-size-fits-all supercomputer is over.
In deep learning, nearly all the work is being done on the GPU, so HPE designed a system with a four-to-one ratio of GPUs to CPUs to improve efficiency, said Scott Misage, general managerof high performance computing at HPE. That’s in contrast to a standard server, which may have a one-to-one ratio — one GPU to one CPU.
HPE is also building special purpose HPC systems for trading applications. The algorithmic code used in match and trade processing is single threaded, “so having a lot of cores on a processor is kind of a waste,” said Misage. “You’d be spending a lot of money and you wouldn’t be using most of the cores.”
The trading process can be made more efficient by using a single-socket server with a chip in this case with eight cores instead of 20. That reduces the system’s overhead, and allows HPE to turn up the clock frequency. That “drives down the amount of time it takes the processor to get to an answer,” said Misage.
Steve Conway, an HPC analyst at IDC, said the special purpose systems mix processor types, communication I/O needs, and the software needed for a particular application.
High performance systems, historically, have been compute-centric, not data-centric. That means the processors were the fastest part of the system, and efficiency was pegged to how well the rest of the system could keep up with the processor, said Conway.
HPE isn’t the only vendor focusing on special purpose systems, but it is the largest HPC vendor. So its actions will be influential on the market, said Conway.
According to the Top 500 count, HPE accounted for 31% of the systems on the list, the largest part of the global pie. Cray was second at 13.8%.
Hewlett-Packard Enterprise (HPE), the market leader in this space, is now producing high performance computing systems for specific needs. The shift is being driven, in part, by the increasing desire for systems that can process data efficiently.
HPE on Monday announced a series of new systems targeted at specific processes such as “deep learning.” This is a branch of machine learning used, in particular, to analyze images and sound.
For deep learning-related processing, HPE built its Apollo 6500 server that can handle up to eight high performance Nvidia GPU cards on a two socket CPU system. The Apollo 6500 is due in the third quarter of this year.
HPE also unveiled the Apollo 4520 System, which is designed to support Lustre implementations, a file system. That system will be available April 18 at a starting price at $8,500. Pricing for the financial services system will be determined by customer requirements.