The problem is that DE Shaw Research made a loosing bet recently when they decided to design custom silicon to solve the protein folding problem while everyone else kept using commodity hardware and worked on improving the software.
I saw a guy I know at a wedding a few weeks ago who works at Shaw Research, and I asked him about this. He didn't give explicit details on custom vs. commodity performance, but he did say they had machines up and running and giving novel results.
I doubt it is fair to say they are ignoring software improvements. I saw Shaw talk about the machine architecture and he had a lot to say about balancing programmability for later software improvements vs. specialized computational resources. Also, it was his algorithmic improvements, the "neutral territory" methods, that were inspiration for the machine. Are NT methods still state of the art for molecular simulation?
How do you know that this is a losing bet? It's still pretty early days. And aren't they hedged with their work on Desmond, which is MD software for commodity clusters?
Custom hardware for chemical physics has been tried multiple times before since the late 70s. I’ve heard most companies only found financial success through bloated government contracts and the custom hardware only provided limited advantages for researchers. In all cases these advantages died out quickly with advances in commodity hardware.
Anton is impressive in that it can provide millisecond long trajectories of protein/solvent systems, but single trajectories are of limited utility. You need many (1000s) of such trajectories for statistical analysis of the molecular system. Further, there already exist several clever methods that leverage chemical statistical mechanics to provide the same analysis without the need for single long trajectories. As an example, see Pande’s work developing Markov models of protein folding using millions of short trajectories between metastable states. This method has already provided a complete statistical analysis of protein folding for proteins that fold on times scales an order of magnitude beyond what Anton can simulate.
As for Desmond, there already exists a plethora of free MD programs (MMTK, LAMMPS, NAMD, CHARMM, Gromacs, and many more). Many of these, especially Gromacs, have already been highly optimized for a range of hardware and I wouldn’t expect Desmond to surpass these free codes by a margin worth dropping dollars.
Personally, I still have high hopes for DE Shaw Research, I just don’t see how their current offerings will turn a significant profit or greatly advance science. I’d love to be proven wrong, and I’m sure they’ll have plenty of additional novel future projects, some of which could be paradigm-shift-changing for chemical physics and molecular biology. My guess is that such advances won’t come from their hardware geniuses, but instead from their math/physics geniuses that will develop new statistical mechanics methods to bend & contract in silico time.
