Supercomputing 2008 kicks off this week, and that means a new update to the Top 500 list.

IBM have given RoadRunner a shot in the arm, boosting sustained Linpack performance to 1.105 petaflops, up from it’s previous best of 1.026.

Cray have done a good job with Jaguar, moving from XT4 frames with a sustained 205 teraflops, to XT5 frames, boosting sustained performance to 1.059 petaflops.

Silicon Graphics come in at third place, with the Pleiades Altix ICE cluster, which they put together for NASA’s Ames Research Center. At 51200 cores and 51TB of memory it’s a bit of a beast, although as a cluster it’s a bit less interesting than a Single System Image (SSI) machine.

SGI‘s Pleiades manages to sustain 487 teraflops, pushing BlueGene/L into fourth place. The previous supercomputer Silicon Graphics built for NASA Ames, Columbia, languishes down in 39th place, which should give some idea of the immense scale of performance improvements taking place.

You can grab the full system stats for Pleiades from http://www.top500.org/system/9832

Silicon Graphics have just announced that more global shared memory is available with fewer CPUs on their Altix 4700 systems. Increased DIMM density now means you can get an Altix 4700 with 2TB of memory, with only 8 processors.

If you’ve got applications that require large amounts of memory but not much in the way of compute-intensive processes, this is very good news indeed.

Global shared memory is memory which is accessible from all processors/cores. So in an SGI Altix with 1024 processors and 4TB of RAM, any one of the 1024 CPUs can access any part of that 4TB of memory. This is due to the design of Silicon Graphics’ large scale systems, which are Single System Image (SSI) machines – all resources are shared.

Clusters work in a different way, where each node has ‘local’ CPU and memory, and this can’t be accessed from another node.

Both SSI and clusters can scale, but in different ways and with different workloads. Shared memory jobs, where you’re doing lots of memory I/O and you can peg your dataset in physical RAM, don’t scale well with clusters, whereas rendering (where discrete jobs can be chopped up and executed in batches) are just right for clusters but not SSI machines.

With lots of memory density enhancements coming down the line, I’m wondering when Silicon Graphics will break through the 4TB system memory barrier?