Supercomputer company Cray Inc. today unveiled what it says is its first “exascale-capable” machine, one that ultimately will be able to do a quintillion calculations per second.

Shasta, as it’s called, is designed for enterprises that need more choice over their compute and networking technologies. It’s also meant to keep things simple by providing a single platform capable of running high-performance computing workloads such as artificial intelligence, data analytics and modeling.

Cray said the idea is to replace the complex heterogeneous cluster systems that generally run most of these workloads today. The company reckons there’s a growing demand for such single systems, as they help to make things more manageable and eliminate performance bottlenecks, so companies can run HPC workloads at scale.

One of the interesting things about Shasta is that users will be able to chop and change the infrastructure running inside it to suit their needs. For example, they’ll be able to choose from processor architectures such as Intel Corp.’s and Advanced Micro Devices Inc.’s customary x86 chips, Nvidia Corp.’s graphics processing units, or even field-programmable gate arrays made by Intel and Xilinx Inc., which can be reprogrammed for specific tasks.

Customers will also be able to choose from different interconnects, which are used to connect the different components of Shasta. These include Intel’s Omni-Path and Mellanox Technologies’ InfiniBand products. A third option includes Cray’s new Slingshot technology, which has been designed for HPC and AI and can deliver five times as much bandwidth as traditional interconnects, the company said.

Cray said Slingshot is compatible with Ethernet and will sport congestion control, adaptive routing and “high-level” quality of service capabilities. With Slingshot, users can build networks with more than 250,000 endpoints, officials said.

“Slingshot’s very low network diameter allows extremely responsive adaptive routing; each switch has a good view of the overall state of the network, so it can make fast, well-informed decisions about optimal paths to take to avoid temporary congestion,” Steve Scott, chief technology officer and senior vice president at Cray, wrote in a blog post. “This allows us to sustain well north of 90 percent utilization, even at large scale, for well-behaved workloads.”

Cray is planning to offer two version of Shasta. The basic version comes with a 19-inch air-based or liquid-cooling-based system that fits within standard data center racks. The more advanced, higher-density version with liquid cooling will run 64 compute blades with multiple processors for each blade. All in all, Shasta is expected to be 10 to 100 times faster than Cray’s most powerful supercomputers currently available.

Although the Shasta supercomputer is targeted at enterprises, the majority of the machines are likely to be deployed in public clouds rather than enterprises’ own data centers, said Holger Mueller, principal analyst at Constellation Research Inc.

“In the past, Cray machines would always run the corporate data center and had the corresponding applications for them,” Mueller said. “But looking at the next generation of quantum computers, it’s more likely that these will be deployed in the public cloud only. Still, it’s not just about the hardware, but also the software. It will be interesting to see what new kinds of applications Cray’s platform can enable.”

Cray said it’s planning to make Shasta commercially available by the end of 2019, so it’ll be some time that’s apparent. However, Cray will be offering a first look at its capabilities during the SC18 supercomputing event taking place in Dallas in November.

Also significant is that Shasta is said to be future-proofed for “exascale computing,” which should be feasible by 2021 or 2022 once the processing technology needed to enable this catches up. Exascale-class computers will be capable of at least one exaFLOPS performance, which is a quintillion or a billion billion calculations per second. That would be a thousandfold increase over the performance of petascale computers, which were introduced in 2008.

Image: Cray