Since 1986 - Covering the Fastest Computers in the World and the People Who Run Them

Language Flags
December 18, 2009

Top 10 Hits and Misses for 2009

Michael Feldman

It was a tough year for HPC vendors  — well, any vendor actually. But in 2009, we also witnessed the acceleration of GPU computing, an increased acceptance of cloud computing for HPC, and the beginning of the post-quad-core era. There were also a number of other interesting developments in the HPC universe. Here are my top 10 picks of hits and misses for 2009.

Hit: NVIDIA Lights a Fire Under GPU Computing

For the past two years, NVIDIA has been steadily building an ecosystem of hardware and software partners around its GPU computing products. With the announcement of the Fermi architecture in September, NVIDIA clearly put itself in the GPGPU driver’s seat. The company’s new GPU for 2010 doubles as a vector processor, adding tons more double precision performance, ECC memory, and support for C++. Multi-petaflop machines with Fermi parts are already on the drawing board. A potential game changer.

Miss: Other Accelerators Hit the Brakes

The rise of the general-purpose GPU parallels the fall of other HPC accelerator wannabes. The high performance Cell processor (PowerXCell), which helped Roadrunner break the petaflop barrier, looks like it will be the last in its line. Apparently, IBM is no longer interested in an HPC Cell sequel. In concert with the PowerXCell demise, Intel has canceled the 2010 introduction of its manycore Larrabee processor. And finally, FP accelerator vendor ClearSpeed is now on life support and is not expected to make a recovery.

Hit: Government Stimulus to the Rescue

Stimulus money designed to blunt the economic recession injected some much-needed funds into supercomputing programs around the world. In the US, the American Recovery and Reinvestment Act of 2009 (ARRA) is funding a range of new work, including a new climate research collaboration NOAA and ORNL ($73.5 million), the 100 Gbps ESnet upgrade for the DOE ($62 million), the Jaguar supercomputer upgrade at ORNL ($19.9 million), and the DOE’s science cloud research project at Lawrence Berkeley and Argonne ($32 million). NASA and the NSF are also parceling out recovery money to HPC projects, generally in much smaller increments though. To date, NASA has spent $86 million, the NSF $92 Million, and DOE $1.58 billion of Recovery Act money, although only a portion of this is being applied to HPC. The good news is that this represents only a small fraction of the stimulus funds that are available to these agencies, so we can expect a lot more money to buoy the government and education segments in the coming years.

Hit and Miss: Cloud Computing in HPC

Along with the aforementioned DOE science cloud research project, a lot of HPC users are experimenting with the cloud concept, but commercial adoption is just getting started. HPC users point to the lack of high performance infrastructure, security concerns, and the cost of making applications cloud-friendly as barriers to adoption. But the prospect of buying HPC cycles and storage on-demand or using a SaaS model to broaden access is too appealing to ignore. Our HPC in the Cloud supplement we ran in November provides a good overview of this nascent delivery model.

Miss: Thinning the HPC Herd

While the HPC market was spared from the worst effects of the brutal economy, a number of vendors in the space were forced to shut their doors, while others were acquired by their healthier brethren. Sun Microsystems, SGI, SiCortex, Woven Systems, RapidMind, Cilk Arts, Interactive Supercomputing, DRC Computer, Quadrics, ClearSpeed, and Verari Systems all became victims of the Great Recession of 2009. Our feature article tells the tale.

Hit: Cray Defies Economic Gravity

In a tough economic climate, Cray has managed to pay off most of the company’s debt, and is looking to generate $285 million in revenue for 2009. The fact that the company is knee-deep in federal government work didn’t hurt of course. Global stimulus money helped fuel some big deals, such as the previously mentioned Jaguar upgrade. Whether Ungaro and company actually posts a profit for the year remains to be seen, but 2009 has essentially been a rebuilding year for the supercomputer maker. A lot of the effort went toward diversification: building out the CX1 deskside lineup and XT mid-range systems, as well as booting up its custom engineering business. Cray’s capture of the number one spot on the TOP500 list with the Jaguar super was largely symbolic, but was a fitting conclusion to a year of progress.

Hit and Miss: A Fickle Japanese Government

Just when it looked the like Project Keisoku would be scrapped, the Japanese government this week agreed to create a special $250 million earmark to continue funding the work. Project Keisoku is Japan’s next-generation supercomputer project that aims to field a 10-petaflop system in 2012 for the country’s science community. The project had its share of trouble in 2009, starting with the withdraw of contractors NEC and Hitachi in April. In November, a government review panel recommended to de-fund the project. With the re-commitment of the government to all this petafloppery, Japan’s science community is breathing a little easier.

Hit: InfiniBand Rising

Despite the commercial debut of many 10GbE offerings this year, InfiniBand adoption in HPC continued to grow in 2009. InfiniBand is deployed in 182 of the TOP500 supercomputers, representing a 28 percent year-over-year increase. Data from InterSect360 and anecdotal evidence points to a dominance of InfiniBand in new HPC deployments. With recent innovations like an MPI offload capability, a memory transfer offload for GPUs, and a performance roadmap that Ethernet can’t match, I suspect InfiniBand’s best days are still ahead of it.

Hit: Intel Nehalem

The debut of the Nehalem EP processor in March marked a milestone for Intel’s server lineup. Because the Nehalem architecture incorporated on-chip memory controllers and the QuickPath Interconnect (QPI), a point-to-point processor communication link comparable to AMD’s HyperTransport link, EP represented the first Intel chip that could go toe-to-toe with the Opterons. In fact, though, since the Nehalem chips were built on 45nm technology, had a larger L3 cache, and supported DDR3 memory, the EP chips outgunned their Opteron counterparts for many applications. AMD gets to counter all that with its upcoming Magny-Cours processor in 2010, but Intel’s next process shrink to 32nm is just around the corner.

Miss: The Joys of High Frequency Trading

In 2009 we learned about the questionable role of high frequency trading (HFT) in the financial markets. A series of mainstream media articles that surfaced in the summer questioned whether these HFT applications, which rely on low latency hardware and sophisticated algorithmic trading software, were doing anything but skimming money from long-term investors. It’s all perfectly legal, extremely profitable, and, thanks to the magic of high performance computing, marvelously efficient.