Sequoia's 20 PFlops

IBM just won the contract for LLNL's next gen flagship machine. The system is basically an upgraded BlueGene/P, with faster CPUs and more cores/socket.

But it's an astonishing level of performance, given that the 100 TFlop barrier was only crossed a few years ago, and the fact they've been able to do it in such a small footprint is pretty amazing too. Of course, it would be fairly straightforward to achieve the same peak performance (not the same sustained), with a bunch of GPUs, and in a smaller footprint (both physical & power), but that machine probably wouldn't have the same level of flexibility, although the per cpu memory bandwidth is probably higher for the GPU system.

For the rest of the industry, this a big deal. It seems that LLNL is now exclusively an IBM house, at least when it comes to Big Iron, and it's hard to see how any other vendor could have any reasonable expectation to win such a contract, despite the open bidding process.