NEC, Appro ink supercomputing pact

Sometimes, when you want to take down a big kill, it takes more than one wolf to do the job. And so it is with specialists in the supercomputing market who are trying to make a stand against the incumbent and much richer general purpose server makers who have the lion's share of sales in the HPC space.

One such specialist, Appro, a privately held seller of blade and rack clusters based in Milpitas, California, has tapped Japanese server giant NEC - no slouch in the HPC arena - to help sell its wares.

Under a partnership announced today, Appro is teaming up with NEC's HPC division to have it sell Appro gear alongside its own SX-9 vector supercomputers, which have a respectable peak processing capacity of 839 teraflops, and its LX Series of X64 clusters that run Linux or Windows and are meshed together using InfiniBand interconnects. (NEC supports both Intel and AMD chips in these LX clusters, while the SX-9 machines use a home-grown vector engine - one of the remaining few left on the market. Cray also still sells vector boxes).

NEC currently sells supercomputers in its home Japanese market, throughout Europe, and in Argentina and Brazil. The company is ramping up its presence in general-purpose X64 and Itanium servers in North America, but in the 1990s, NEC was accused by Cray of dumping its supercomputers on the U.S. market and was eventually slapped with huge import duties after Cray made a big political stink about it.

The upshot is that Appro doesn't have to worry too much about NEC invading its home turf with X64 clusters, but then again, Cray has a different set of senators (from Washington rather than Minnesota) this time around, and they are probably more software protectionists than the senators from Minnesota were hardware protectionists a decade ago.

Appro has been on the cutting-edge of InfiniBand clusters, and just this past June, it was the first vendor to demonstrate 40 GB/sec Mellanox interconnections running inside its Xtreme-X1 clusters, which are based on quad-core Xeon server nodes. The company's Xtreme-X2 nodes are based on Advanced Micro Devices' Opteron processors.

Both machines are based on a super-dense, second-generation blade server design, a kicker to the HyperBlade design that put Appro on the map a few years ago. The Xtreme-X series scales to 6 teraflops of number-crunching power in a single 44U rack, and using quad-core chips, can put 512 cores supported by 4 TB of main memory in that rack.

Appro has a reference architecture that shows it can span up to 442 teraflops with 36,864 cores, but it has yet to sell such a big box. It has, however, made some pretty big wins at Lawrence Livermore National Laboratory, Los Alamos National Laboratory, and Sandia National Labs in the States, which are three of the biggest consumers of teraflops in the world. (It takes a lot of oomph to simulate nukes so you can get around the Nuclear Test Ban treaty and redesign warheads). Significantly, Appro's clever and energy-efficient designs gave it a win at the University of Tsukuba in Japan. NEC obviously noticed this.

By the way, this is not the first time that NEC has partnered with what would seem to be a competitor. The biggest supercomputer deal that Sun Microsystems did in a long time was the "Tsubame" cluster installed at the Tokyo Institute of Technology (abbreviated TiTech), which was comprised of 648 X4600 Opteron servers equipped with ClearSpeed math accelerators, all linked together with a giant InfiniBand switch and rated at 47.4 teraflops. The gear is currently ranked 24 on the Top 500 supers list. The way that Sun won the deal was by making NEC its implementation and services partner. ®