Named for microbiology pioneer Anton van Leeuwenhoek, Anton is currently being built with 512 highly specialized processors. These are clocked at just 400MHz, and the machine has modest memory, but its architecture lets it process problems in a massively-parallel way. Ultimately, that'll offer a performance boost of 1000x over current complex molecular simulations. And that's great news: these bits of math are how drug design works. It's different to processing done by existing supercomputers like BlueGene/L in that it will look at molecular behavior over a longer interval. That means scientists could discover new biological processes. "If you can do 1,000 times longer, real proteins come into play" as team leader David Shaw puts it. Anton should be in operation later this year. [ACM Library via NYTimes]
@pharago: I would say "less like".
DESRC has spent a lot of money hiring a vertically integrated team, capable of designing custom architectures, programming environments, and novel algorithms capable of exploiting both of the former, for the purpose of reaping remarkable speedup on those problems that they care about. In that this architecture is specialized to use massive parallelism to impact a specialized application, it is similar to the path that GPUs have taken over the past decade. However, the architecture here is tailored specifically for the problem of performing MD simulations (even down to a particular MD implementation), including computational primitives and a memory architecture that NVIDIA/ATI won't be putting on a GPU anytime soon.