MOST OF HPC HAPPENS UNDER THE RADAR
January 31, 2020 Michael Feldman
While supercomputers are arguably the most exciting segment of the high performance computing market, the majority of systems deployed in any given year do not fit into this elite category. As a consequence, most of the annual HPC server spending goes to systems that cost less than $500,000 – about $8.4 billion of the $13.7 billion annual total, according to Hyperion Research’s numbers for 2018. That represents an attractive opportunity for system integrators and resellers to sell bread-and-butter clusters to a much broader set of users.
Of course, a fair number of HPC users have the wherewithal to configure, purchase, and install these systems on their own, but the further you travel down the food chain, the less likely you are to find that expertise in-house. Just as important, deployments of these small and medium-sized HPC tend to be done in a much more streamlined fashion than, say, a $100 million supercomputer that’s been in the planning stage for three years or more. Even in the US federal government, where familiarity with the ins and outs of configuring and installing HPC machinery is at the highest level, these smaller systems are purchased more like off-the-shelf commodities. Again, these smaller systems are widely deployed; they just don’t garner the notoriety of the big-name supercomputers at the Department of Energy and Department of Defense. You will rarely encounter a press release for such machines.
Which brings us to Koi Computers, which specializes in selling HPC systems and services to US federal agencies. Launched in 1995, the Chicago-based company started out by offering generic computer hardware to state and local agencies. According to Catherine Ho, Koi’s federal business development manager, that changed around 2000, when Koi got a contract supplying HPC hardware to the Department of Energy. That’s when its HPC business took off, she says.
Although Koi still sells commodity off the shelf (COTS) hardware, like desktops, to federal customers, that’s a sideline now. HPC was identified as a growth segment that conveniently overlapped with Koi’s expertise in supplying computer equipment to the feds, and to a lesser extent, the higher education market. The sweet spot for them is anything from a tens of nodes up to a couple hundred, all based on off-the-shelf components. Some of Koi’s most valuable expertise has nothing to do with technology: the company is vetted to work with a number of federal contract vehicles, including the National Institutes of Health Information Technology Acquisition and Assessment Center’s (NITAAC) CIO-CS and NASA’s SEWP, as well as the GSA vehicle IT Schedule 70. The company is also well-acquainted with the other processes involved with navigating the federal bureaucracy.
While familiarity with government paperwork might seem like a rather mundane kind of expertise, this capability helps these agencies streamline their planning and procurements, which greases the bureaucratic wheels. That, in turn, speeds up system deployments while also potentially lowering costs.
According to Ms. Ho, these contract vehicles – contract templates really – are used across a wide array of US government agencies. Customer include the usual HPC suspects like the DOE, DoD, and NOAA, but also less well-known players like the National Institutes of Health (NIH), the Federal Aviation Administration (FAA), and the Internal Revenue Service (IRS), among others.
As you might suspect, Koi suppliers are a who’s who of HPC providers: Hewlett Packard Enterprise, Dell, Supermicro, Mellanox Technologies, Nvidia, Intel, AMD, Tyan, and Gigabyte, to name a few. When the company works with partners like HPE and Dell, it’s basically acting as a reseller; when working with Supermicro and Tyan, it slides into the system integrator role.
Customers generally come to Koi with specific performance and budget requirements, along with an overview of what they hope to accomplish with the system. According to Ms. Ho, they usually have some idea of the intended applications for the system. In some cases, the customer will specify hardware preferences as well. Using that input, Koi can go to its hardware partners and come up with a set of possible system options and configurations for the customer to choose from. The company generally provides system installation and benchmarking, but also more specialized services like acceptance testing. Support and maintenance services are additional options.
Business is good, Ms. Ho told us, but couldn’t share specific numbers. (Koi is a privately held company.) Generally though, they are seeing the kind of growth reflected in the overall expansion of the HPC market. Apparently, the interest in AI-powered HPC systems has filtered down into these smaller systems as well, reflected in an uptick in the number of GPU-powered clusters being purchased by federal agencies. To deal with the Koi’s expanding prospects, the company recently built a larger, state-of-the-art facility in Illinois to be able to build and test larger clusters and greater numbers of them. Ms. Ho said the larger power capacity available at the new facility will enable engineers to benchmark these systems more efficiently.
Koi is a good reminder of how much HPC is taking place below the radar. Even though we’re on the cusp of the exascale era, it worth considering that most of the HPC systems being deployed today still operate at terascale levels.