The engineers have done their job. All that’s left is that annoying detail of selling the product. So thought Liquid Computing, a Canadian tech startup that builds Linux-based high performance computing systems. But so far, the potential of the technology has exceeded Liquid’s ability to capitalize on it.

The company introduced its LiquidIQ system offering in October 2006. The architecture offers a computing model that uses AMD Opteron processors and a tightly integrated interconnect fabric to achieve high levels of energy efficiency and scalability. The company is initially going after HPC customers, with a particular focus on federal government and oil & gas HPC users. More recently, they have targeted service providers, including managed services providers, hosting providers, and Software as a Service (SaaS) solutions providers. Despite favorable reviews of their LiquidIQ offering from analysts, the company is still looking to gain traction in its intended markets.

“As an emerging technology company, one of the biggest hurdles that we’ve had to overcome is brand recognition, says Tom Kreidler, Vice President of Federal Sales, Liquid Computing, ”Essentially, people were asking: ‘Who the heck are you guys?'”

Another hindrance to the process may have been the small sales force. Until recently Kreidler could hold the sales team meetings in any room with a mirror in it. He joined the company in January, coming from Juniper Networks as the vice president and general manager of the federal systems division. Since then, Liquid has added Nick Weston to lead the company’s energy sales team, and Michael Bohlig to pursue the service provider market. All three are veteran Sun alumni and bring with them the customer Rolodex for their respective application areas.

In April 2007, Liquid announced their first customer deployment: Network Computing Services, Inc., the systems integrator and computing facilities manager for the Army High Performance Computing Research Center (AHPCRC). In May, the company announced a partnership with Dalhousie University’s Faculty of Computer Science to help research and develop parallel On-line Analytical Processing (OLAP) on Very Large Databases (VLDB). These two field deployments represent the only LiquidIQ installations where users are not just kicking the tires.

Kreidler admits that the company had expected a faster ramp-up. So they retrenched and starting listening more closely to their prospective customers. What they found was a reluctance to invest in Liquid products until the company could prove they were going to be around for awhile. Some of this just amounts to demonstrating you have the resources to weather the startup cycle. Since February 2005, Liquid has collected over $43 million in financing, which theoretically should give them plenty of time to get their sales pipeline going. But company revenue and purchase orders are the real proof point for prospective customers.

“Sometimes when people say ‘I don’t know if you’re going to be here in a year,’ at least in the back of your mind, you’re tempted to say: ‘Well some of that’s up to you,'” jokes Kreidler.

Once customers get beyond the longevity issue, they start to ask what makes the Liquid technology different from the rest of the pack. On the surface, LiquidIQ is comparable to high-end blade x86-based systems from Sun, HP or IBM, or in more scaled out configurations, a Cray XT-type supercomputer. However, Kreidler notes that comparing their offering to a typical blade system is not quite fair. Since LiquidIQ incorporates the interconnect fabric into the system itself, you’d have to add a router and interconnect switching functionality to the blades to get a side-by-side comparison. The high level of communication fabric integration is the basis of Liquid’s differentiation when compared against more conventional blade servers.

The company has put together some benchmark results, which puts the Liquid product well into supercomputing performance territory. In May, Liquid released results of two HPC Challenge (HPCC) benchmarks in which LiquidIQ outperformed the Cray XT3 on both the ping pong latency metric (1/3 the latency of the XT3) and the random ring and natural rings metrics (1/2 the latency of the XT3). Kreidler says this is achieved at about a third of the cost of an XT3. Liquid also claimed a STREAM benchmark record for four-socket Opteron processor-based servers. Using Opteron 8220 processors, LiquidIQ achieved more than 20,000 megabytes per second on all of the STREAM benchmark bandwidth measurements. They also claim to be getting similarly impressive results when running real user applications in some of their target domains, like oil and gas.

Because of the high performance interconnect, the architecture excels at taking advantage of the second processor on dual-core processors. While not achieving true linear performance, which is limited by shared memory, according to Kreidler the LiquidIQ modules are achieving 1.8x – to 1.9x performance increases for dual-core systems compared to single core systems; their competitors are getting 1.2x. If true, this would give Liquid a significant opportunity to offer a lower cost solution — for both up-front costs and operating expenses — for a given performance level. If the company can maintain this level of scalability as multicore processors move to quad-core and beyond, this advantage would become even more interesting.

Some of the sales and marketing efforts may be starting to pay off. Last week, they received a purchase order from one of the big federal government integrators (as yet unnamed) with plans to deploy a system in an embedded environment. The hardware will be used to support high performance computing in a military vehicle application. The on-board liquid cooling inherent in LiquidIQ made it particularly suitable for this type of environment. The company also has a number of high-visibility users who are engaged in field trials of LiquidIQ. The company expects purchase orders from one or more of these installations in the near future.

“Our guess is that we’ve got a bunch of users who are waiting to see who’s going to jump first,” says Kreidler.