5832 The Sicortex Newsletter of High Processor Count Computing

Green is an attitude, not a deprivation. Green is about getting it right at the beginning, not flinching at the end. In computer systems, green is as much about cool software as it is about cool hardware.

More than anything, green is about products that get out of their own way so their inherent efficiency and power can see the light of day. SiCortex is showing that, in computer systems, green performance is the most valuable performance of all.

Read more or download a pdf version.

Journeys may be sequential, but petaflops are not. Peta–flops are insistently parallel. The only way to get there is to let a thousand processors zoom. Or perhaps a hundred thousand. And the only way to do that efficiently is to develop and tune the codes on thousands of processors. Taking a code from the sequential computing era and spawning it a hundred thousand times is almost never the answer. New ways of thinking and cost-effective, high-processor-count, development environments are mandatory. SiCortex is uniquely positioned to support this new wave of peta-innovation.

Read more or click to download a pdf version of the newsletter (3 MB).

The first thing we see when we look at a SiCortex SC5832 or SC648 is the array of 64-bit Linux processors. It is all too easy to think of them as "doing all the work." But the work of modern computing involves much more than just arithmetic. It involves constant interchange of information, which is where the systems' arrays of autonomous DMA Engines come in. Sharing the node chips with the Linux cores, they operate independently. and when the true work of an HPC application is tallied up at the end of the day, they have often done more than half of it.

Read more or click to download a pdf version of the newsletter (2.1 MB).

FabriCache™ is SiCortex’s unique way of accessing reserved portions of main memory as though it is a parallel disk file, but at hundreds of GBytes/sec. Every processor has access to every address of the file in a fully-coherent way, without any special programming, because all references are managed the same way that disk files are managed. Whether it is comparing genomes or staging time-step data, FabriCache embodies the SiCortex philosophy of performance through simplicity.

Read more or click to download a pdf version of the newsletter (2.1 MB).

Traditional cluster nodes are made up of chips from multiple vendors. Each may have ways to read out its performance characteristics, but the formats are incompatible and the results can be incomprehensible. Because the SiCortex system nodes were designed from a clean sheet of silicon, performance instrumentation is complete, consistent, and integrated. On a SC5832 or SC648 you can take the pulse of your application at each stage of its execution. And you should.

Read more or click to download a pdf version. of the newsletter (2.1 MB)

Getting the heat out of the SiCortex SC5832 and SC648 unleashed a virtuous circle of benefits that is especially sweet for communications. Having all the processors on a single backplane not only restored performance and reduced cost, it also opened topology options unthinkable in the age of cabled clusters. How did the SiCortex architects get their heads around these new options? They used computer visualization. The ultimate choice, a Kautz graph implemented on a 48-layer copper backplane, is not only fast, it is also quite pretty.
Read more or click to download a pdf version of the newsletter (2.3 MB).