That contract included about $100m to build the Blue Gene/L ridiculously parallel supercomputer (well, what else do you call that which is far beyond massively parallel?), with the remaining $190m going into the ASCI Purple machine.
The initial ASCI Purple specs called for the cluster to be comprised of 196 64-way Squadron servers using dual-core Power5 processors running in the vicinity of 2 GHz, with a total of 12,544 processor cores and using a daisy chain of IBM’s Federation High Performance Switch (HPS) system interconnection switches to link them all together.
The ASCI Purple machine has 12,288 Power5 processors running at 1.9 GHz, which is comprised of 1,536 of p5 575s. The expected peak Linpack performance rating for ASCI Purple a few weeks ago, before the trials were run, was around 93.4 teraflops.
Supercomputers like ASCI Purple are usually built in stages, starting with then-current technology and a subset of the hardware and gradually building up to the fully capable machine. You build in stages, do performance and architecture tests, and as the machines pass milestones, they move on to the next stage.
Last Friday, IBM announced that it had finished building the pieces of the ASCI Purple machine and that the boxes will be delivered to the new Terascale Simulation Facility at LLNL in August, with final acceptance testing to begin in October. That facility has some pretty impressive power requirements, since it needs to juice up those 12,288 processors, 50 TB of main memory, and a 2 PB (petabytes, or 1,000 TB) file system with over 8,000 disk drives. The whole shebang takes 4.8 megawatts of power, which is about as much electricity as 4,800 homes use, and generates 16 million BTU/hour in heat. Getting rid of that heat also a huge amount of electricity.
To get to the stage where LLNL would take delivery, IBM powered up a demonstration system of 1,280 servers, which ran a mix of supercomputer applications modeling turbulent hydrodynamics and unstructured mesh radiation transport.
The deal required IBM to measure theoretical peak performance on this system such that peak performance on one workload plus sustained performance on the other added up to 101 teraflops. As it turns out, the system came in at 111 teraflops, which obviously made IBM pretty happy. Big Blue also ran the Linpack Fortran benchmark on the system with 1,185 nodes active, and it came in at 60.5 teraflops.
The ASCI program was founded in 1995 to encourage computer makers to hit the 100 teraflops performance point in the 2004-2005 timeframe, but this goal is important not in the abstract. It had a very specific purpose aside from generating innovation in computing, and that was to run massive simulations to make sure that the arsenal of nuclear weapons in the US that have been sitting in holes in the ground for decades still work.
