The title of their press release is “Liquid Computing throws down the gauntlet with HPC Challenge results,” subtitled “LiquidIQ beats Cray XT3’s posted latency performance.” They go on to cite these results:
HPCC ping pong minimum latency of 0.46 microseconds and a maximum latency of 2.58 microseconds. Liquid’s HPCC ping pong latency is one third of the latency posted by Cray XT3 on the HPCC benchmark (Reference HPCC data: http://icl.cs.utk.edu/hpcc/hpcc_results.cgi).
HPCC Random Ring and Natural Rings latencies are measured at 3.63 and 3.00 microseconds, respectively. Liquid’s latencies are less than half of the latencies posted by Cray XT3 for the same tests.
They are clearly focused hard on “beating” Cray. This doesn’t make any sense to me. This company isn’t positioned significantly in the scientific market right now, and at any rate has technology that is way too new to claim can be successfully deployed at the scale that Cray and the other “old timers” can. In fact, their own “about” text at the foot of the press release says the following:
Liquid Computing Inc. is first to deliver a new class of computer system called LiquidIQ™ to meet the needs of scalable computing users within Enterprise High Performance Computing, xSP and Telecom markets.
While I’m sure Cray would sell to these guys, none of these markets are primary targets for Cray today and probably weren’t in the design criteria for the XT3. Now, they seem to have good technology, and may well a vendor of large-scale HPTC solutions one day. But they aren’t there yet, and so the comparison seems a little vacuous.
True, Cray is the the elder statesman of the supercomputing companies and has an incredible legacy. But we are long since past the days of Cray’s dominance of the market. Today we have a vibrant HPC market with offerings that meet a variety of needs. No one offeror is “the” choice anymore and swinging this hard at Cray smells a little desperate.
Then there’s the matter of their specific results. Yes, 3-ish microseconds is a fairly good time. But there are seven systems, including two XD1s from Cray (no longer available; fair enough), with sub 3 microsecond times, and 4, including SGI and NEC, with 3.x microsecond times. What about those? Not worth a mention?
Also, they only quoted two performance figures: what about the other seven? The whole point of the HPCC is to provide information to the community that illuminates a broader range of a machine’s performance dimensions. If random ring latency isn’t important to me, I don’t guess I care how the Liquid machine stacks up to the XT3.
Overall this looks like a rookie mistake by a new company trying to make a name and a market for itself. Rookie mistakes are easy to make, and I certainly won’t hold it against them.
But I’d suggest they take a moment, breathe deeply, focus on the technology, deploy solid systems, and let their market grow. They have a long time, and a lot to prove, before they need to start tearing into other companies to support their own growth.