HighPerformanceComputing
Latest
Intel's next many-core chip will be a true stand-alone processor
Intel's current Xeon Phi doesn't really fulfill the promise of many-core computing -- it's a co-processor that needs a 'real' CPU to function. That will change when the next-generation Knights Landing model arrives, Intel revealed at the Supercomputing Conference this week. The 14-nanometer chip will be available as a stand-alone model that can run all software, like a traditional processor; since it won't have to shuttle data between two components, it should be faster, easier to program and cheaper, too. There will also be high-speed memory built into the chip, as well as a number of (unspecified) architectural tweaks. Knights Landing isn't likely to ship until late 2014 or 2015, but it could be worth the wait for researchers, server operators and anyone else who wants massively parallel computing power.
Intel opens up about its 'Knights Corner' supercomputer co-processor
HotChips is the show where chip makers come to show off their latest slices of silicon, and Knights Corner architect George Chrysos spilled the beans on Santa Clara's Xeon Phi co-processor. The unit's designed to bolt onto Xeon chips to help supercomputers crunch the numbers faster, by handling the "highly parallel" grunt work necessary for genetic and climate modeling, among other things. Chrysos has lofty goals for the hardware, hoping that it'll contribute to "scientific and technical progress," while we're just excited to see if it can help the company reclaim its Top 500 crown from IBM.
IEEE pushes for Ethernet standard between 400Gbps and 1Tbps, hopes to head off big data crunch
Ethernet might seem passé to those of us toting Ultrabooks, but it's important enough to provoke a crisis for internet providers and many of those who depend on high-speed computing networks for a living: based on the rises of streaming video and social networking, the IEEE is worried that many of those large-scale networks will need 10Tbps of total bandwidth just to avoid a logjam in 2020. To that end, the standards body has formed a Higher-Speed Ethernet Consensus group that's mulling a new, breakneck-speed format reaching either 400Gbps or 1Tbps, depending on whose approach you'd favor. Fight the urge to pick the 1Tbps option on instinct, however. Both options would depend on bonding multiple connections together, and the faster of the two formats could lead to some expensive and very ungainly cables if it's not handled well. A meeting is scheduled for late September in Geneva to at least begin hashing out the details. Although we won't be wiring our homes with terabit Ethernet anytime soon, the standard should come quickly enough that the Googles and Netflixes of the world can satisfy our data addictions for a good while longer. [Image credit: Justin Marty, Flickr]
Intel slips details of Poulson-based Itanium 9500 in advance, teases a big boost to 64-bit servers
If you think Intel took awhile to roll out the Xeon E5, imagine the mindset of Itanium server operators -- they haven't had any kind of update to the IA-64 chip design since February 2010, and they're still waiting. Much to their relief, Intel just dropped a big hint that the next-generation, Poulson-based Itanium is getting close. Both a reference manual and a Product Change Notification have signaled that the new, 32-nanometer part will get the Itanium 9500 name as well as a heap of extra improvements that haven't been detailed until now. We knew of the eight processing cores, but the inadvertent revelation also confirms about a 50 percent hike in the interconnect speed and a matching increase in the cache size to 32MB. Clock speeds also start where current Tukwila-running Itaniums stop, with four processors between 1.73GHz and 2.53GHz giving the line a much-needed shot of adrenaline. Few of us end users will ever directly benefit when Poulson ships to company server farms later this year; after these increases, though, don't be shocked when the database at work is suddenly much quicker on its toes.
Intel christens its 'Many Integrated Core' products Xeon Phi, eyes exascale milestone
Been wondering when the next big leap in high performance computing would hit? Well, Intel would like you to believe the time is now and the name of that revolution is the Xeon Phi. Formerly codenamed Knights Corner, the Many Integrated Core product is pushing the field of supercomputers into the era of the exaflop by squeezing a teraflop of performance into a package small enough to plug into a PCIe slot. The Phi brand will, at first at least, be applied to specialized coprocessors designed for highly parallel tasks. The chips are built using Intel's 22nm manufacturing process and 3-D TriGate transistors, piling in more that 50 cores in an effort to combat the inroads made by GPU companies like NVIDIA in the supercomputing space. For more info check out the presentation (PDF) and blog post at the source links.
NVIDIA outs a pair of Tesla GPUs to electrify your supercomputer
NVIDIA's announced a pair of Tesla GPUs that'll give some extra pep to your supercomputing tasks. The K10 and K20 units harness the power of Kepler to add more muscle to the company's scientific and technical computing arm that supplies gear to the Barcelona Supercomputing Center and Tokyo's Tsubame 2.0. Internal tests reveal that the hardware is around three times faster than the company's Fermi GPUs -- with the latter card expected to arrive at the end of the year. The company didn't announce pricing, since its aiming them squarely at the big academic institutions, defense contractors and oil explorers -- but if your surname is Buffet or Abramovitch, then they might sell you one at trade.
Intel's gettin' kinda heavy, it's got the power, gonna break your heart
Intel spent most of yesterday standing in the bar at the SC11 conference showing off the size of its computing prowess. Admittedly, it's on a roll; the Xeon E5 processor powers ten of the Top 500 supercomputers in the world, and Intel chips as a whole are found inside 85 percent of all the machines on the list -- not to mention the E5's newly minted PCI Express 3.0 compatibility. The Santa Clara chip maker is also building a new Exascale lab at the Barcelona Computing Center, a ten petaflop "Stampede" machine at the Texas Advanced Computing center and several other machines for Government agencies like NASA. The Seattle shindig was a great excuse for Intel's Rajeeb Hazra to show off the new Knights Corner co-processor, built with a new Tri-Gate 3D 22nm process that packs 50 cores into one strip of silicon. In fact, the only thing Intel wasn't bragging about was being bested by the record breaking Fujitsu K, the mere mention of which caused everyone to go a little bit too quiet and begin to glower (we assume).
NVIDIA unveils second-gen Tesla GPU-based workstation cards
NVIDIA's Tesla GPU-based high-performance computing workstations and add-in cards have been on the market for a whole year now, and to celebrate, they're getting birthday cake, balloons, and an upgrade to GT200-based chipsets. Like AMD's recently-announced FireStream 9250, the new T10P processing units are capable of breaking the teraflop barrier, up from the first gen's paltry 518 GFlops, and they're up to 240 cores from the first gen's 128. You'll have to shell out to get all that horsepower, though: the entry-level, 900GFlops C1060 PCI card will sell for $1699, while the four-GPU 1U S1070 blade will sell for $7995 for two PCIe-interface version or $8295 for the single PCIe connect model. The standalone Tesla workstation has been discontinued, as customers were increasingly buying the cards, so it looks like those are really fast collectors' items for now. So, who's going to be the first to add one of these bad boys to the Engadget Folding@Home team?[Via Tom's Hardware, thanks Matan]
NVIDIA launches Tesla: GPUs are the new CPUs
We've seen a couple cautious attempts at leveraging the raw floating-point capabilities of modern high-powered graphics cards, but NVIDIA is taking the gloves off with the launch of Tesla, its new general-purpose computing platform built on the 8-series graphics cards we all know and love. According to NVIDIA, the only way to skirt the inevitable collapse of Moore's Law is to join the GPU and CPU together, so two of the three Tesla configs are in the form of workstation upgrades -- a $1,499 single GPU PCI Express card and a $7,500 dual-GPU "deskside supercomputer" that plugs into a custom PCI controller. The truly crazy can pony up a full $12,000 for NVIDIA's first rack units, the four-GPU Tesla S870, which has a peak performance of 2 Teraflops. We're hearing the card and deskside unit will be available in August and that the servers will start shipping in November or December -- perfect for the Engadget Folding@Home holiday rush.
8-core NextDimension PC stretches the definition of portable
No, the laptop industry hasn't bypassed the idea of quad-core laptops in favour of octo-core machines, since NextComputing's 8-core flextops are machines you wouldn't want anywhere near your lap. Based as they are around the Intel Xeon 5300 processor, the NextDimension Pro and Evo can take advantage of the relatively low power requirements of 100W for two quad-core CPUs (down from 160 Watts for its power hungrier desktop equivalent). For a sense of just how expandable the NextDimension machines are, consider that they can hold up to twelve 160GB 2.5-inch 7200RPM hard disk drives. They also manage to pack in four PCI / PCI-e slots in the Evo model, and Firewire, Gigabit Ethernet, and support for 24GB of memory through four DMA channels on both models. These 20-pound desktop machines with a handle will be shown off at Interop Las Vegas later this month, but beyond that we're not sure when or for how much you'll be able to lug one away.
