[ruslan]

Seems like those guys knew nothing about HPCs. Why didn't they run LINPACK test? It's essential to measure any parallel computing system even of two cores. Also, any first grade CS student knows that the most significant part of an HPC is not the cores, but the network. You need to connect hosts using Infiniband or alike. Using regular ethernet is futile because of high latency, you will waste 90% of CPU cycles in data exchange/syncronization wait loops. I bet they could achieve a way better results on just 1/3 number of cores or even less.

[mukyu]

LINPACK is about as useful of a benchmark as BogoMips.

""". Genentech benefited from the high number of cores because its calculations were "embarrassingly parallel," with no communication between nodes, so performance stats "scaled linearly with the number of cores," Corn said."""

That is a direct quote from the article.

[awj]

Not that I think it entirely bursts your internet tough guy rant, but amazon does offer [cluster computing instances](http://aws.amazon.com/ec2/hpc-applications/) for exactly this purpose. Granted, they only have 10 Gigabit ethernet, but it's not exactly like this is some failure cluster running all over a busy datacenter on 10M ether.