Pocket cluster for $ 99

    The Adapteva company (which you most likely hear about for the first time) plans to make a supercomputer that will be available to everyone . Since 2008, they have been developing energy-efficient RISC processors for orders from manufacturers of smartphones and other mobile devices.

    “We are going down the food chain,” says CEO and founder Andreas Olofsson. But Adapteva wants to give its technology directly to people through a Kickstarter project if they raise at least $ 750K with an ultimate goal of $ 3M.


    Adapteva calls his project “Parallella: A Supercomputer For Everyone”, which is a 16-core processor with a total frequency of all 16GHz cores, 26 GFLOPS and $ 99. If they reach the goal of $ 3M, they will make a motherboard with a 64-core processor (45GHz and 90 GFLOPS) for $ 199. (Adapteva considers frequency and performance as the sum over all processor cores). Both boards also include a dual-core ARM A9 SoC with a 16- or 64-core RISC processor, which acts as a coprocessor to the main one. Adapteva claims that they have achieved energy efficiency of 70 GFLOPS per watt and 25 GHz per watt respectively.

    The promised delivery time (subject to a collection of $ 750K) is May 2013 for the 16-core version.

    The device will be a fully working computer with Ubuntu 11.10 ARM, 1GB RAM, two USB 2.0 ports, 16GB MicroSD card, HDMI and Gigabit Ethernet. The open source SDK will support C, C ++ and OpenCL. The device will be approximately 8.5 x 5.5 cm (the size of a Raspberry Pi).

    There are great similarities between Parallella and projects like the Raspberry Pi (a Linux computer for $ 35 and an Arduino for $ 30), but Olofsson claims that Parallella will be 10 to 50 times faster than the Raspberry Pi (compared to the 64-core version) and while only 3 times more expensive. He notes that with a price of $ 99, it is much cheaper than most current parallel computing platforms. Boards containing Adapteva processors sold by their partners now cost several thousand dollars.

    As a scope, Olofsson calls the development of mobile and embedded systems, the creation of new programming languages, the study of parallel algorithms, etc. “What people can do with Parallella is unlimited,” he says. "We hope that they will be used by open source project enthusiasts who today lack this type of platform to meet their needs."

    In addition to enthusiasts and developers, Olofsson hopes that parallel computing will be widely distributed among ordinary firms. “Today, there is a gap between researchers and ordinary business,” he says. “I know that in the end, parallel computing will become ubiquitous, it's just a matter of time, but I would prefer it to happen now, rather than three years later.”

    The definition of "supercomputer" seems somewhat vague. The slowest supercomputer in the top-500 gives 61 TFLOPS. A cluster of hundreds of 16-core Parallellas will cost $ 10,000 and provide 10 TFLOPS, he said. Even if it is not a supercomputer, then it can be useful for so many people.

    Adapteva says they need to raise $ 3M through Kickstarter to begin production of the 64-core version of Parallella. While a 16-core chip is manufactured using a 65-nanometer process, a 64-core chip will be produced using a more complex and expensive 28-nanometer process.

    Kickstarter recently began to refuse hardware developers, saying: "Kickstarter is not a store." But Adapteva managed to convince them that they were not ordinary retailers.

    But why did Adapteva even go on Kickstarter? The company raised $ 2.5 million in venture capital, but Olofsson says it is "very small for an electronics developer. ... Our research budget is probably 1/1000 of this at Intel. To implement this project, we need millions of dollars. We spoke with venture capitalists, but for them the model of the “iron” startup does not work anymore. They do not get a good return on investment. ”

    Despite the fact that Parallella is based on existing processors, fundraising is the only way to get a price of $ 99, and they come from the cost of production at the Global Foundries factory. And such a low price is key to the success of the project, says Olofsson.

    Discussion on slashdot

    Now a few words from myself:

    If they do what they want, then this will truly be a breakthrough in entering the wide market of multiprocessor systems. Unlike GPGPUs (NVidia CUDA, etc.), which are mostly SIMD, this thing is really MIMD and each core can execute completely independent code from other kernels. I think that with the advent of google Project Glass there will be a strong need for such pieces of iron for example for image processing in the implementation of augmented reality.

    Yes, while they lose in performance to video cards, but at the same time their chip eats 2 (!) Watts. Characteristics of one processor node (assembled from different parts of the documentation):

    • 32-bit single-precision floating point only (no double-precision)
    • Core local store global-addressable on 32-bit flat address space; can R / W outside address space with proper address mapping support
    • 32KB local store on each core, 32 GB / s @ 1 GHz
    • External interface 2 GB / s * 4 (four directions) = max 8 GB / s
    • Fetch instruction from global address space
    • Fast inter-core write, slow read requests (1 for each 8 cycles)

    But their main difference is that their architecture, toolchains, libraries and APIs are completely OpenSource (hello NVIdia with your binary drivers and architecture for NDA).

    Well, even if we ignore this chip, then IMHO in itself a development board with an ARM A9 for $ 99 is more than normal.

    Survey results on a kickstarter regarding possible areas of application

    On the question of whether it is much weaker than a video card:
    In the single-chip version, yes, but the authors added MINI-CLOUD ($ 495) and CLUSTER ($ 975) options. For this price, they promise a solution of four / eight boards ( 72/144 processors (namely the processor, not the cores, as I understand it)), a gigabit switch and a power source.

    Also popular now: