U.S. supercomputer sets speed record

Ya, but I heard Visa slowed it to a crawl.

I don't quite believe these numbers. LLL itself says that their expanded BG/L tops out at a mere 433 teraflops - well below the one petaflop cited in the article and substantially below records held by a number of IBM's newer P-series machines.

An IBM Blue Gene/P, for example, has four cores (4 ea PPC440+FPU compute engines) per node as opposed to two cores/node found in the L-Series machines. There are a number of benefits to packing more cores onto the same piece of silicon (an ASIC, in this case): For one thing, the fact that the cores share the same real estate means much shorter prop delays for communications between engines. Not only, but the quad-core parts are clocked at higher rates than their two-core, L-Series siblings. The P-Series' nodes also have larger L1/L2 caches plus more and faster DDR RAM. Topping it off is the fact that compilers can optimize a larger class of problems efficiently into parallel 4-processor groups. Not so easily done on an L-Series with its paltry two cores per node.

<gets down from pulpit>

I looked on LLL's website but found no reference whatever to their L-Series machine breaking the petaflop barrier, but would someone else here please also check out the facts? I, for one, don't believe the article's numbers.

Eu

The article incorrectly states that the machine is a Blue-Gene/L, as suspected. But neither is it a BG/P. This machine is a new development from IBM, and you can read IBM's very own press release (and not some oblivious technoob-journalist's vague impressions) here.

I want one.

"Dear Santa..."

You smell that? That smell! That SkyNet smell! Nuthin else in the world smells like that.

Smells like - Judgement Day.