greenbiz.com— Researchers from Goethe University and the Karlsruhe Institute of Technology (KIT) have used relatively cheap components to create a vastly more efficient method of processing data.
Apr 6, 2010View in Crawl 4
The alternative, in a system like this, is probably enterprise class fibre channel drives and whatnot. Those things aren't actually much cheaper than SSDs.Pretty sure a $50 WD Caviar wasn't going to provide the throughput they were looking for.
ARMs have pretty poor floating point performance (one of the reasons why the GSM audio codec is optimized for integer math.) The Atom includes some SIMD functionality, which can increase some FP operations quite a bit, especially when you hand-craft your code to take advantage of it.
I would like to see them dump the polygons we have come to associate with GPU's and focus on raytracing. A highly parallel rendering method which is much superior to fudging triangles...
Yeah dude, you don't know what you're talking about. Most super computers just crunch sparse matrices like dutchguilder says. They have a very "streamable" workload. And yes, cpus are more suitable for transaction processing but the compute workloads on even the biggest commercial transaction systems are small compared to scientific supercomputer workloads. (however, IO workload is much greater in commercial tasks).
Sorry, SSD does not belongs in the same sentence with cheap. Maybe in 10 years, but definitely not now.Also, once you include in additional labour and space needed to house a clustered low end CPU, I highly doubt it's worth it.
dutchguilder2Apr 6, 2010
448 cores producing 500 Gigaflops while sucking only 200W is amazing. Only a few years ago a decent size reservoir simulation project might run 24x7 for a week on a beefy PC; the Tesla C2050 will probably complete that same run in just 15 minutes! <a class="user" href="http://www.nvidia.com/object/product_tesla_C2050_C2070_us.html" rel="nofollow">http://www.nvidia.com/object/product_tesla_C2050_C ...</a>
explodingzebrasApr 7, 2010
how much longer do these machines have to run to complete the same task as faster CPUs?
merrebornApr 7, 2010
Performance per kWh is a pretty common, real world computing metric.Performance per dollar is up there too. This addresses both.
merrebornApr 7, 2010
The alternative, in a system like this, is probably enterprise class fibre channel drives and whatnot. Those things aren't actually much cheaper than SSDs.Pretty sure a $50 WD Caviar wasn't going to provide the throughput they were looking for.
jbmcbApr 7, 2010
ARMs have pretty poor floating point performance (one of the reasons why the GSM audio codec is optimized for integer math.) The Atom includes some SIMD functionality, which can increase some FP operations quite a bit, especially when you hand-craft your code to take advantage of it.
myztryApr 7, 2010
I would like to see them dump the polygons we have come to associate with GPU's and focus on raytracing. A highly parallel rendering method which is much superior to fudging triangles...
h0dgesApr 7, 2010
Maybe, then, I have a problem with the word 'blazing'. Perhaps 'green' would have sufficed? ;)
zephyrprimeApr 7, 2010
Yeah dude, you don't know what you're talking about. Most super computers just crunch sparse matrices like dutchguilder says. They have a very "streamable" workload. And yes, cpus are more suitable for transaction processing but the compute workloads on even the biggest commercial transaction systems are small compared to scientific supercomputer workloads. (however, IO workload is much greater in commercial tasks).
sndreamApr 7, 2010
Sorry, SSD does not belongs in the same sentence with cheap. Maybe in 10 years, but definitely not now.Also, once you include in additional labour and space needed to house a clustered low end CPU, I highly doubt it's worth it.
asforonedayApr 7, 2010
You do realize that literally translates to "I am like turtles", as in, "I am similar to turtles", right?
lnmagicApr 10, 2010
Electricity adds up over time. Each Watt you use in a machine that's on for a full year costs about $1.