Since exiting the commodity server market in 2014, IBM Corp. has been able to pay much more attention to its high-end machines. The latest fruit of the vendor’s efforts is a trio of Power LC models that are built to run complex artificial intelligence and analytic applications.

The most noteworthy of the systems is the two-socket S822LC, which is based on a new flavor of Big Blue’s flagship POWER8 processor that was developed with the help of partners from its open-source chip consortium. What sets the design apart is its built-in support for NVLink, an interconnect created by Nvidia Corp. to speed up communications between its Pascal-series GPUs and CPUs from third party vendors. The technology allows data to flow 5-12 times faster across the chipset than the PCIe interfaces historically used for the task.

IBM used NVLink to pair the two POWER8 processors in the S822LC with four accelerators from Nvidia’s Pascal P100 series, which can provide as much as three times more throughput than the older M40 and K40 models. Big Blue claims that the combination of the technologies makes its new server much better suited to run high-performance computing workloads like deep learning software than x86 machines. In one internal benchmark test, the machine was shown to provide 80 percent more performance per dollar than comparable 2-socket systems based on Intel Corp.’s E5-2600v4 Broadwell chips.

IBM says that the S822LC’s speed has already led several enterprises and government agencies to make pre-orders. Among them are the U.S. Department of Energy’s Oak Ridge National Laboratory and Lawrence Livermore National Laboratory, which have separately commissioned the vendor to build them new supercomputers using Nvidia’s NVLink technology. They will employ the server to see how their scientific applications run and make optimizations before the supercomputers come online next year.

Meanwhile, organizations with more mundane requirements can choose between the two other Power LP systems that Big Blue unveiled today, which use standard POWER8 chips with PCIe interfaces. The first is called the S822LC and can be ordered with two K80 accelerators plus up to 97 terabytes of flash or disk. Its sibling in turn is the S821LC, which only supports one Nvidia chip and 32 terabytes of storage but requires just half the space. Both models are based on a two-socket design geared towards analytic applications that process large quantities of information.

Image via Geralt