At Supercomputing ’08, SGI has announced “Molecule” – a research project that could put 10,000 cores in a single rack. According to Shahin Kahn:

Analysis of applications and market trends shows that several customer environments can potentially benefit from the combination of significantly higher density, memory bandwidth, memory bandwidth per instruction, and performance per watt

Well, duh. Welcome to the party, guys. Some people did that analysis years ago, and have been shipping products based on exactly those ideas. No “research project” about that, no “could” do it a couple of years from now; just products that customers can buy today. It’s not 10K cores in a rack – it’s 5832 in approximately a double rack – but that actually includes things that SGI didn’t include in their calculations. Good luck putting your 10K processors and memory for them in a single rack. Oh, and what good is a drawer full of energy-efficient processors with massive memory bandwidth if they can’t talk to one another? Now you’ll have to eat some of that space and destroy some of that power efficiency with a nice fat Infiniband switch or ten, and you’ll still have lousy communications compared to your processing power and memory size/bandwidth. Lots of processors sitting around waiting for messages means longer job completion times, and time has to be factored into energy-efficiency metrics to be meaningful.

Hitching this idea to the “Atom” brand is good marketing, but is it good engineering?