Here at, when we write about assembling computing clusters in the cloud, it tends to be on a grand scale. Think endeavors in high-performance computing (HPC) on Amazon’s Elastic Compute Cloud (EC2) like in 2013 when a chemistry professor and software company Cycle Computing assembled 156,314 cores for an 18-hour run that reached a theoretical speed of 1.21 petaflops. Or when simulation software firm Schrödinger rented 50,000 cores on EC2 in 2012 for $4829 per hour.
But sometimes just several dozen cores from a computing fairy-godmother will do. That was the case for an undergraduate Hyperloop team from the University of California, Irvine (UCI). The team assembled in 2015 to compete in a series of contests sponsored by SpaceX after the company’s CEO, Elon Musk, drew up a whitepaper envisioning a super-fast form of transportation that ran on magnetic skis in a low-pressure tube. After UCI’s team, called HyperXite, won a technical excellence award in an early 2016 design competition, the team had to actually build the thing in time for the January 2017 pod contest at SpaceX’s headquarters in LA. This meant a lot of computer modeling would have to be done.
Around that time Nima Mohseni, a mechanical and aerospace engineering undergraduate at UCI, joined HyperXite as the team’s simulation lead. He toldon a phone call that HyperXite was originally looking at doing its modeling on a 24-core system owned by the University—a system that limited students to 72 hours of time per project. But HyperXite was sponsored by Microsoft (among others), and the software giant intervened to introduce the undergraduates to the people at Cycle Computing.
Cycle Computing takes computing capacity from cloud services and provides its clients—often research teams and private companies—with software to enable large-scale simulations in everything from risk modeling for life insurance disbursements to testing potential solar cell materials. For HyperXite, Cycle Computing helped assemble a 128-core cluster on Microsoft Azure to support all the necessary data and compute workload. To do the actual modeling, HyperXite used simulation software ANSYS running the Azure instances, which each had 16 Intel Xeon 2.4Ghz cores and 7GB of RAM per core.
This allowed the team to model the aerodynamics of the pod as well as its braking system. Just after the 2016 design competition, HyperXite decided to revamp the entire shape of the pod because their original design had been too heavy. The team decided to change the shell of the pod from aluminum to carbon composites, but before building a first draft of this shell on a real pod, Mohseni ran simulations to make sure the change would allow the pod to keep its structural integrity. “Right when we changed one part, something else had to change,” Mohseni told. With the new shell, some parts of the pod were too weak, and would break upon the pod’s braking. Mohseni said the team built several iterations that broke in real life too, but having the benefit of a lot of computing capacity allowed the team to iterate faster and throw out clearly unworkable shell shapes.
“If we did not have access to the computational servers, we’d have to run smaller simulations and not get as accurate a result,” Mohseni told. “That’s the big difference.” Having more than five times the computational capacity cut work that would have taken weeks down to days, and allowed the team to model its pod in far more detail.
Jason Stowe, CEO of Cycle Computing, toldthat HyperXite’s experience isn’t unusual outside of academia. “What we’ve noticed, and the thing that Nima mentioned, was that by getting answers back quicker, people tend to benefit from that in two ways,” Stowe said. “First, you’re accelerating the time to answer… By being able to get answers back far quicker, you can do more iterations in less turnaround time. Second, you can say ‘great, we’re done’ and get [your project] done faster.” Stowe noted that Cycle Computing helped HyperXite pair ANSYS with the Azure cores, but if UCI had had an internal cluster it wanted to add computing resources too, the HPC software company could have helped it do that too.
Ultimately, the HyperXite team took its pod through the preliminary tests at the SpaceX Hyperloop competition, but it didn’t get a chance to run its pod in the near-vacuum environment due to time limitations. (Actually, only three of the 27 teams competing got vacuum track time because the length of three-quarters mile track had to be depressurized and repressurized between each run—a process that took an hour, or sometimes more.)
“It more of an educational experience,” Mohseni said, “it was [about] getting a whole team together of over thirty, forty students,” to teach them how to become engineers. And with sponsors donating an awful lot of computing resources, HyperXite was able to do just that.