WSJ reports on the launch of SiCortex, a company I’m involved with. Talk about a contrarian bet–to back a “supercomputer” company in the age of Linux clusters and to spend non-trivial amounts of capital on developing sophisticated chips, interconnects and enclosures when everyone else gets their stuff made in China for hundreds of dollars per machine.
It all began with the right question. Rather than focusing on how to increase performance, the founders asked themselves how they can reduce power consumption. That changed everything. The typical approach in the PC space is to push for more performance, which has the unfortunate side effect of making machines run very hot, which effectively puts a limit on performance because chips needs to be far enough apart so that they don’t melt and you just can’t pack enough of them in a data center before having to do unnatural acts with power and cooling. SiCortex machines run cool, which enables densely packed chips, which leads to low latency and high parallelism, which significantly increases delivered performance.
The company’s top-of-the-line supercomputer has 972 such chips — or 5,832 processors in all — in a cabinet a little less than six feet tall. A cluster with comparable performance might need eight to 10 cabinets, and would draw about 10 times more power than SiCortex’s machine.
So, just another fancy supercomputer then? Not really. WSJ should more accurately describe SiCortex machines as “superclusters” because they look like traditional Linux clusters that communicate using MPI. Contrary to traditional supercomputers which had custom architectures and required scarce brainiacs to parallelize algorithms to these specific architecture, SiCortex machines can run industry standard codes. I guess I learned Occam in school for nothing.
BTW, there is a cheaper $200,000 machine with only 648 processors for you power-hungry geeks with a successful startup experience under your belt.