In high performance computing, SuperMicro is a force that is seen but rarely heard. They are often found at all the major supercomputing events even if it’s uncommon to hear about large procurements or installations. However, don’t mistake this lack of news as absence of action—the company’s role in HPC is strong and is set to grow, according to CEO Charles Liang.
In a recent conversation, he told us that while the OEM-driven business model and the lack of publicity around their big wins at large-scale centers can lead to the sense that they’re a sideline participant in HPC, this is a misperception. For instance, he says that although they’re not listed as a system vendor for a single machine on the Top 500 on the most recent listing, they estimate between 15-25% of all Top 500 machines are built with their components—most likely motherboards, at the very least. This is a quiet, but notable achievement for a company whose HPC business makes up only 25% of overall revenue.
With that in mind, and considering that 95% of the company’s business is in the server and storage arena, they topped out the last quarter alone at $428.1 million. This marked a 4.5% jump in quarter-to-quarter revenue—taking them to record revenue heights with full-year figures in the $1.47 billion territory. That success is being plunged back into the company this year, with new operations and R&D centers, including one in San Jose, not far from their headquarters. To fill that space, it’s clear they’re on a hiring spree as well.
The new facility, which can be considered an “extension” of 36-acres with a 300,000 square-feet building to complement their nearby main location, will address research and development goals, but also will be dedicated to logistics and system integration. Liang said the new campus, dubbed the Green Computing Park, should significantly help the company grow its business in the U.S., which will be mirrored by similar efforts in Taiwan and the Netherlands to serve those regions. “In the last three years we’ve doubled our engineering efforts and in that time have tripled our facility size,” said Liang.
He expects Super Micro’s position in HPC to improve further this year with the addition of a new line that balances the two concerns they’re hearing most from supercomputing’s end users—to offer a balance between high efficiency and high performance—all with an eye on budget. Accordingly, the company has rolled out a new family of machines along the “Ultra” line, which will include the UltraRack and UltraTwin when they’re released. Liang says the targets are creating a balance of efficiency and performance in a dense framework. “We have removed the traditional I/O rigidity and bottlenecks by developing a set of new multifunction I/O modules. This allow us to have a much higher performance on barriers, networking and storage solutions,” he noted.
The new Ultra architecture series most notably supports the Xeon E7-2880 v2 processors and offers 32x DIMMs per node for data-intensive applications. As Liang noted, “This new server is part of our innovative Ultra Architecture featuring high-density, performance and cooling optimized serverboards supporting Intel’s highest core-count processors, massive memory footprints and the latest high-performance storage, expansion and network technologies. UltraTwin is the first in a line of upcoming servers to take full advantage of this groundbreaking architecture and expands our MP solution range with support for Intel Xeon E7 v2 series processors and peripheral technologies.”
Like other server makers, SuperMicro is holding out for the new Haswell chips, but is seeing solid growth with existing lines, including the FatTwin, which Liang says has seen remarkable growth. This HPC-oriented line saw a 150% jump in revenue based on last year’s figures. Further, GPU-based systems are also on a tear at Supermicro, with 75% higher growth year over year and 13% higher sequentially.
Tau Leng, VP and GM of Super Micro, who also joined our conversation, said that they’re selling a diverse array of machines in HPC from small 16-32 node clusters to those with many thousands of nodes. In addition to topping the Green 500 list, they’ve had some significant installs at Pacific Northwest National Laboratory in the last year, among others. The PNNL machine, which was put together by Atipa, ThEMSL HPCS-4A is a 42x 42U rack s with 1,440 compute nodes and 2,880 Intel Xeon Phis, granting it 3.38 petaflops at peak with an addition 2.7 petabytes of storage.
“We have invested a great amount of engineering effort to perfect our Twin server technology and now offer an unrivaled range of server solutions optimized for practically any scale application. With our new 4U 2-node FatTwin featuring dual Xeon CPUs and six Xeon Phi coprocessors per node, science, research and engineering programs can increase and accelerate project deliverables with maximized utilization of budget, resources and space,” said Liang.
He added, “Since we were founded 21 years ago, we’ve concentrated on the building block approach. Some applications need more memory, some more storage or bandwidth, some efficient cooling. We will continue in this direction, putting our work into these blocks.”