It looks like a group of researchers at UC Berkeley have come up with a rather unique way of solving the problem of getting supercomputers past the processing power / energy consumption barrier, with them now touting the possibility of using millions of low-power embedded microprocessors instead of conventional server processors. That tantalizing prospect has apparently already lead to a deal with Tensilica Inc, which will provide the Berkeley researchers with some of its Xtensa LX extensible processor cores to use as the "basic building blocks in a massively parallel system design." Ultimately, the researchers say they could one day build a massive supercomputer consisting of 20 million embedded microprocessors at a cost of $75 million, which they say would have a power consumption of less than 4 megawatts and a peak performance of 200 petaflops. That, they say, would be enough for it to create climate models at 1-kilometer scale or, as the researchers put it, more than 1,000 times more powerful than anything available today.

[Via TG Daily]

Well, it sure looks like IBM is keeping all its supercomputing bases covered these days, with it not only working on a chip-sized supercomputer, but a global-scale shared computer that'd be capable of "hosting the entire internet as an application." That latter word comes in the form of a white paper ambitiously dubbed "Project Kittyhawk" (we're guessing they found "Project Mulitvac" a little too obvious), which aims to explore the construction and implications of such a massive scale computer. That apparently wouldn't be a SETI or Folding@home-style shared computer consisting of everyday PCs, however, with it instead relying on IBM's petaflopping Blue Gene/P as a common platform, which would effectively be able to run any web-scale application you could throw at it. Of course, none of that has moved very far beyond the page just yet, so you can rest easy that there's still no supercomputer out there that's capable of bringing the entire internet to the halt on a whim, at least that we know of.

[Via Slashdot]

Off hand, we can't think of how we'd truly utilize the horsepower generated by a 72-processor cluster shoved into a "whisper-quiet, low-power deskside cabinet," but we'd happily draw up a plan if forced. SiCortex -- the same folks who delivered the bicycle-powered supercomputer -- has introduced its new SO072 Catapult, which features a standard Linux environment, 48GB of RAM and a trio of (optional) PCIExpress slots. This aptly categorized high performance computer (HPC) sucks down less than 200-watts of power, sports a pair of gigabit Ethernet ports and has room for six internal hard drives. Reportedly, each of the 12 SC072 nodes is a multi-core chip with six CPU cores, and while $15,000 may seem steep for your average tower, we'd say this is a pretty good value considering the hardware.

[Via Gadgetopia]

The TOP500 supercomputing list was just announced and IBM's BlueGene/L system has kept its crown. In fact, IBM's and the Department of Energy's co-developed monster at Lawrence Livermore has occupied the number 1 position since 2004. Of course, an upgrade was required boost the Linpack benchmark to 478.2 TFlop/s from the 280.6TFlop/s the machine was clocking just 6 months ago. The top 10 swath is dominated by the US, Sweden, and Germany with India breaking into the list for the first time at the number 4 position with its HP Cluster Platform 3000 BL460c system measuring 102.8TFlop/s.

[Via Impress]

We've certainly seen PlayStation 3s used in academia before, but a "significant gift" by IBM to the University of Maryland-Baltimore County is enabling the institution to further develop its Multicore Computational Center. The facility now utilizes 32 Cell processors, and researchers at the lab are getting set to "develop applications" to harness all that power. Reportedly, the supercomputer will dabble in "visualization and computer modeling applications for aerospace / defense, financial services, healthcare / life sciences industries and weather." Hit the link below for a video report.

[Thanks, Ryan O.]
Read - IBM, UMBC partnership
Read - Video report

In an attempt to illustrate the low power requirements of its latest Linux-based supercomputer, a company called SiCortex put together a demonstration showing the massive number-cruncher performing billions of calculations per-second while a team of between eight to ten bicyclists pedaled to keep it going. The NextFest presentation was a ploy to showcase the small power consumption of the new machine, and to point out that energy conservation for both system power and cooling are major issues facing manufacturers and end-users. As the custom Trek bikes churned out 260-watts of power each, Dr. John Mucci -- the company's CEO -- said, "Ten years ago, this analysis was impossible on even the biggest computers, and now it's being done on a bicycle powered machine," then added, "Ten years from now, we'll probably be doing it right down at the docks as they unload the ships." Which sounds kind of weird, but isn't.

[Via I4U]

Just months after scientists were able to run a quantum computer simulation on an everyday PC, we're now hearing that a Calvin College student / professor tandem have created an inexpensive, portable supercomputer for crunching massive chunks of data on the go (and on the cheap). Dubbed Microwulf, the wee beast is hailed as a "machine that is among the smallest and least expensive supercomputers on the planet," and when not being checked as baggage on a flight, can reportedly process 26.25 gigaflops of data per second. The system itself touts "four dual-core motherboards connected by an eight-port gigabit Ethernet switch," and when initially constructed, it cost just $2,470 to build. Talk about a solid price-to-performance ratio.

[Via Slashdot]

We've seen what (little) a quantum computer can do, but a pair of curious scientists flipped the equation around and sent a humdrum PC to do a supercomputer's work. Professor Peter Drummond and Dr. Piotr Deuar were able to "successfully simulate a collision of two laser beams from an atom laser using an everyday desktop computer," which would typically only be attempted on a substantially more powerful machine. Notably, the achievement wasn't entirely without flaw, as the purported randomness in the testing eventually "swamped everything" and forced the simulation to be halted in order to gather any useful data whatsoever. Unfortunately, we're all left to wonder exactly what kind of machine was used to chew through such grueling calculations (Compubeaver, perhaps?), but feel free to throw out your suggestions below.

[Via Physorg, image courtesy of ACQAO]

It was but two years ago that we all sat in awe of IBM's beautiful 280.6-teraflop masterpiece, and it's finally safe to call that workhorse a thing of the past. Today, Big Blue announced the second-generation of its famed supercomputer (dubbed Blue Gene/P), which reportedly comes close to "tripling the performance of its predecessor." In an attempt to showcase its might, IBM suggested that 27 million clinical trials could be computed "in just one afternoon using a sliver of its full power." Additionally, the engineers went the extra mile by crafting a highly scalable system that comes in "at least seven times more energy efficient than any other supercomputer," but we're sure those 294,912 processors (each touting four PowerPC 450 CPUs) can churn through some energy when utilizing the whole petaflop. And just think, Sun probably thought the 500 or so teraflop production from its forthcoming Ranger would be the (momentary) top dog.

Read - IBM's Blue Gene/P, via CNET
Read - Sun's Ranger, via Wired

Efforts to model the human brain (on IBM's Blue Gene, ironically) haven't reached the point of finality just yet, but it looks like the supercomputer has already tackled a smaller, albeit similar task at the University of Nevada. The research team, which collaborated with gurus from the IBM Almaden Research Lab, have ran a "cortical simulator that was as big and as complex as half of a mouse's brain on the BlueGene L," and considering that it took about 8,000 neurons and 6,3000 synapses into consideration without totally crashing, it remains a fairly impressive achievement. Notably, the process was so intensive that it was only ran for ten seconds at a speed "ten times slower than real-time," and while the team is already looking forward to speeding things up and taking the whole mind into account, it was noted that the simulation (expectedly) "lacked some structures seen in an actual brain." Now, if only these guys could figure out how to mimic the brain and offer up external storage to aid our failing memories.

[Thanks, Richard L.]

While Field Programmable Gate Arrays (FPGA) haven't benefited from a good deal of buzz just yet, things could be taking a turn, as both Xilinx and Altera caught a few eyes at IDF. We've already seen the unique, albeit highly specific chips in a supercomputer, but the dedicated coprocessor / accelerator modules could be landing beside your Intel Xeon CPU. Essentially, the devices plug "directly into the processor socket of dual- or quad-socket servers" in order to provide "high performance application acceleration ranging from 10x to 100x compared to processors alone, while simultaneously reducing overall system power consumption." The modules act as targeted CPUs, effectively computing very specific tasks in a much more efficient fashion than a general microprocessor can alone, which could boost the speed of scientific, financial, and life science applications that rely on very particular calculations. Of course, mainstream adoption still has quite a ways to go, but the quicker we get dedicated physics and AI coprocessors to go along with these snazzy new GPUs, the happier (and poorer) we'll be.

[Via RobotSkirts]
Read - Altera demonstrates FPGA at IDF
Read - Xilinx demonstrates FPGA at IDF