It seems that well over half of the high performance community we speak with still has no idea how they ended up with a career in supercomputing specifically. Some filtered in from end user experience, others’ software development roads forked into Fortran…the stories go on…
Despite the throng that stumbled into supercomputing from across the disciplinary divides, there are indeed plenty of others who chose HPC early on—and keep coming back to it, despite a world rife with broader options. For people like this, including newly-minted CTO at Cray, Steve Scott, there’s really no place like home.
During his time away from Cray following an almost 20 year career, Scott served as CTO for the NVIDIA Tesla division followed by a relatively short career at Google where he served as Principal Engineer, toiling away on the type of stripped-down hyperscale systems that could, quite understandably, make an HPC hardware junkie homesick. But of course, that homesickness is just conjecture.
“Google was great, but I came to realize over time that my heart was in HPC. I love the impact that HPC has on all kinds of important problems in science and society,” Scott said following the announcement of the new appointment.
Some could argue that Cray could benefit from a strong home team that is seasoned, refreshed, and ready for a new path toward broader roads. Despite their penetration into new markets and growth, Cray has experienced some leadership changes over the last year. The CTO slot opened earlier in the summer with the departure of Bill Blake as well as their YarcData division head, Arvind Parthasarathi, both of whom left to chase entrepreneurial efforts. While they’ve picked up other notables to fill the void, including a new board member from the big data side, Max Schireson, CEO of MongoDB, bringing Steve Scott back to head the technical vision for the company firms up their focus as an engineering company—and could even hint at their coming investments in the very things Scott is known for on the interconnect front in particular.
Now, in addition to his time at Google and NVIDIA, Scott can walk back into an HPC business that’s doing its best to push the big data message to a wider set of potential end users. While he told us that the infrastructure “lessons learned” at Google are far different than would apply to supercomputing, he has a broader appreciation for what the big data trend really means—and where traditional HPC systems and software fit into that picture.
“There’s a definite convergence underway,” Scott said of the meshing between the data-driven enterprise world and large-scale high performance computing. He said that while he’s a “hardware guy” first and foremost, Cray has had to track the analytics and data-intensive business by bolstering its software investments. “We’re a software engineering company,” he agreed, “but more important, we’re a systems company.”
When asked about how he plans to take Cray, a company he’s known since his early days fresh out of college at Cray Research, into the more mainstream high-end enterprise big data world, he noted that there is a virtuous cycle—one that Cray is uniquely positioned to take advantage of. “With more data, there is a need for HPC and with more HPC, there will be constantly be new tools developed to keep pace. That in turn creates more sophisticated models, which create more data and that cycle continues.”
On the topic of coming full circle, both personally and technically from the Cray perspective, he says we can expect HPC systems to complement what’s happening in big data and vice versa since there are common outstanding problems to be solved, particularly in terms of data movement.
“If we think about what makes a good scientific computer, in large part, it’s not really about the FLOPS or the compute, it’s about moving the data—even in classical high performance computing—the quality of the system, the thing that makes it a real supercomputer is how it can move the data. When you take that same strength and start to apply it to statistical systems like Hadoop style or graph analytics the performance you get is largely determined by your ability to move data more quickly.”
For someone like Steve Scott, data movement means something a bit different than it might to the more mainstream big data crowd. His involvement in a fine-grained approach to moving data—the mighty interconnect—was a defining point in his career at Cray and, as some assumed, at NVIDIA where he drove the GPU leader’s Tesla business through the introduction of NVLINK and other key technologies.
“We see our traditional scientific supercomputers being good for data analytics,” he said. So you can expect to see platforms that can and need to do both because we’ll see more workflows that need to be capable of both types of workloads.
We expect that Peter Ungaro is finding this change refreshing—perhaps even more so than this.