teraflops
Latest
Fastest-ever Windows HPC cluster nets 68.5 teraflops
In the grand scheme of supercomputers, a homegrown cluster constructed by Microsoft is just a boy among men. However, said rig has set at least one record by becoming the "fastest-ever Windows HPC cluster." At the International Supercomputing conference in Dresden, Germany, it was announced that this beast ranked 23rd in the world "with a problem-solving performance of 68.5 teraflops." The National Center for Supercomputing Applications utilized a beta version of Windows HPC Server 2008 to hit the aforesaid mark, and if you're curious as to what it took to get there, try 9,472 cores of processing power. There's more where this came from for the hardcore nerds in attendance right in the read link.[Via Slashdot]
IBM pushing vertical stacking in next wave of supercomputers
Next-generation cooling technology isn't the only thing IBM's R&D crew is spending time with, as the chip giant has recently made plans to hit up "vertical stacking technology" in order to make the next wave of supercomputers really crank. Supposedly, "laying chips vertically -- as opposed to side by side -- reduces the distance data has to travel by 1,000 times, making the chips faster and more efficient." The new format will place chips directly atop one another and connect them with "tungsten filled pipes etched through the silicon," which will subsequently eliminate the need for wires and increase the speed at which data can flow. The questionably-dubbed "3D chips" will reportedly operate around 40-percent more efficiently than existing renditions, and considering that Intel is purportedly cooking up a similar agenda in their own labs, that "end of 2007" release date is quite likely to be accurate.
Japanese supercomputer breaks the petaflop barrier
Japan, which once topped the list of world's fastest supercomputers with NEC's Earth Simulator, has seen its position deteriorate in recent years in the face of faster machines like IBM's 280-teraflop BlueGene/L. Well now it looks like a new Japanese rig is poised to regain the top of the charts, and the most amazing thing about Riken's MDGrape-3 -- besides its claimed 1 petaflop performance -- is the fact that it cost only $9 million to build, giving it a per-gigaflop pricetag of just $15 (compared to the $140/gigaflop cost of IBM's top dog). Developed in conjunction with Hitachi, Intel, and NEC subsidiary SGI Japan, MDGrape-3 is being tasked with helping the pharmaceutical industry model new drugs, as it can calculate the chemical bonding properties of a proposed drug-protein combo in mere seconds. While BlueGene/L contains a whopping 130,000 processors distributed over 65,000 nodes, Riken's closet-sized machine needs only 4,808 chips to achieve four times its speed for certain applications. Oh, and despite the impressive-sounding performance, due to the specialized nature of its design, its unlikely that you'll see MDGrape-3 rocking a game of Doom anytime soon.[Via Slashdot]
