Earlier today Cray announced the XC30-AC (air-cooled) supercomputer, a pared-down version of its top-of-the-line XC30 system, aimed at the technical enterprise market space. The news was timed to coincide with the 2013 Cray User Group (CUG) meeting in Napa Valley, California.
The Cray XC30-AC Air-Cooled Supercomputer
Where the XC30-LC (liquid-cooled) product is targeted at Fortune 100 companies, the air-cooled version is intended for the rest of the Top 1000, notes Jay Gould, Cray’s senior product marketing, in an interview with HPCwire. These “mid-market” systems were previously designated as m-versions.
What’s different about the XC30 series, says Gould, “is we can scale up to massive machines like we always have, but now we can also configure down to more aggressively priced machines by customizing the packaging and introducing cost-savings. Previously, the -m was an attempt to size down the machine, people would talk about mid-range, and minis, etc.
“Technical enterprise is what we’re shooting at, and we don’t want anyone who coughs up $500 thousand to $3 million for a Cray to think that they got a diminutive mini version of something or a neutered version of a supercomputer.”
To Cray, “technical enterprise” encompasses pricing, performance and applications. The segment ranges roughly from $500,000 systems to $3 million systems. Above that line lies Cray’s high-end systems. The AC systems will initially extend from 20-200 teraflops, a window that will expand along with Moore’s Law-timed processor iterations.
The AC naming convention follows in the footsteps of Cray’s rebranded Appro systems, but Gould was clear that the AC system leverages all the technology innovation and investment of the XC30 series, optimized for the technical enterprise space.
“We’re economizing all this high-end innovation that we developed for our flagship line and finding ways to make it more aggressively priceable for smaller markets, smaller datacenters and a new class of users within the existing spaces that we already play,” he adds.
Built on the pillars of the XC30 architecture, the AC system uses the same processor technology, the same compute board, and the same processor daughter cards. It is this “Adaptive Supercomputing” strategy that Gould says allows Cray to turn on a new customer trend or industry movement or support new devices without having to redesign the architecture from scratch. Cray has already pre-announced its support for the Intel Phi and NVIDIA GPUs in XC30 and that IP will port straight to the air-cooled version as well. The I/O blades are the same, as is the HPC-optimized Aries interconnect, but the Dragonfly network technology holds an important distinction.
Dragonfly was designed with three ranks. The first rank combines blades within a single chassis via a backplane, while the second rank connects local chassis to each other via passive copper electrical cables. The third rank – a network built of active optical cables that provides row-to-row communication – is intended only for the most massive supercomputers. For smaller systems, like the XC30-AC, this very expensive technology is overkill.
“So while there is this difference in the network topology, it’s still the same architecture so everything is compatible going up. The same software, same software stack, same partners, same ISV vendors, and same middleware vendors are all in play. The Cray ecosystem remains in tact,” says Gould, “as does the Cray reliability, service and support.”
In addition to employing a less expensive network implementation, the Cray XC30-AC offering is distinguished by its economized packaging, cooling and power options. Each cabinet is self-sufficient: a single high-efficiency fan sucks air in from the bottom and blows it vertically through the cabinet and out the top. There’s no need for liquid coolants or plumbing infrastructure, which is what allows Cray to target new customer types. There’s also no requirement for raised floor datacenters, in fact these systems could even run on cement slabs in a garage.
What ultimately makes this configuration possible is that the cabinets are smaller and less densely populated. The flagship systems are stuffed with blades, as many as will fit, which necessitates a powerful cooling system. In this setup there are 16 vertical compute blades per cabinet, relying on a single fan for cooling. Because the cabinets are self-sufficient, up to eight cabinets can be joined without the need for additional cooling support. To accommodate smaller datacenters or server rooms, the XC30-AC offers a lower-power 208-volt option in addition to the 480-volt standard.
Cray designed the AC version to meet the needs of a new classes of users. There’s a big move to leverage modeling and simulation across multiple verticals. In manufacturing, energy, finance and the pharmaceutical industry, businesses are looking to transition from physical to virtual prototyping to improve ROI and boost time to market. Whether it’s designing the perfect golf club head or developing the world’s most sophisticated turbines, users want to be able to simulate those things from the bottom up rather than building multiple prototypes and dealing with lengthy development cycles.
Cray guides new customers through the selection process by sifting through all of their requirements to understand their application requirements and use model.
“Sometimes they just know at a high level what they’re shooting for,” says Gould. “In other cases, they’ve got a really specific request for information or request for bid where they itemize everything because they’ve done this a lot.
“Some of the new classes of users that we’re talking about haven’t necessarily used high-performance computing before so they don’t even know all of the questions to ask when learning about a system. So when we go through the engagement, we find out whether they have a cement slab floor in their computer room or a raised floor with air flow everywhere, whether they have liquid plumbing or not, so we can guide them based on their performance requirements and their budget to the right solution for their application.”
Not every organization can operate a hundred million or three hundred million or billion dollar datacenter, says Gould. “Some of these new customers don’t even call it a datacenter. They may call it a computer room, computer lab, or server room.”
Additionally, not every computing requirement can be addressed with a cluster. Clusters are a nice fit for capacity type applications, a use case that Cray affirmed when it acquired Appro. But as Gould points out, the supercomputing vendor is seeing new demands from existing customers and from interested prospects that can only be addressed by a capability-optimized computing platform.
The scaled-down XC-30 should appeal to customers who get time on large Cray machines at national labs. While the advantages of leadership-class systems like Titan or Blue Waters are undeniable, the allocation process has the downside of long wait times and other constraints. The AC racks will allow customers to own their own machine and be 100 percent in control of their schedule and time to market, and they can still utilize the big machines for larger-scale workloads.
Gould stresses that there is complete compatibility from two to 200 cabinets and beyond, ultimately using the same software, the same IP, and same kind of networking.
“It’s going to be a compatible migration, not starting from scratch and porting your applications all over again,” he notes.
This level of compatibility was no accident, as Gould explains:
“We went into this whole portfolio over the last several years open-minded, knowing that we wanted to do a high-end version and a more frugal technical-enterprise version,” he says. “Instead of building the world’s biggest, fastest supercomputer and then trying to figure out how to cut it into pieces, we built it from the ground-up so we could configure single cabinets with air-cooling all the way up to the world’s biggest machines – 480-some-odd cabinets – and be able to be flexible enough to change the networking for the bigger machines and scale down the networking for the smaller machines. That took a lot of time and investment and that was one of the biggest challenges: ‘how do you use one technology for everything?’ and I think we hit this very well.”
The product line is available now and is already shipping. Early customers include a global consumer electronics company and a global financial services company, highlighting the move to non-traditional HPC segments. Cray wants to do its part to ensure that innovation is not limited to the top 100 companies. There is a lot of room for growth and there are many Fortune 1000 companies with an emerging need for a class of supercomputers that fits within their datacenters and their budgets.
“In the macro view,” says Gould. “HPC [spending] is still going up, and the region we are targeting – the half-million to three million dollar price-point – is actually a growth area, not regressing or shrinking, and this is part of our strategic plan to continue to target the right applications and the right integrated systems for those markets.”