Well, here's the thing: if you're going to go DIY, you've got to go all the way! Think outside of the box... necessity is commonly known as the mother of inventions. Ergo, if you want a DIY cluster, you've gotta make the case or do an overhaul to something else big, like this practical example!! And in this case, it uses multiple power supplies... but you can DIY the main power supply cables, as explained here. But, if you don't want to do it yourself, there is only two things that will drive the Multi-(mini-)ATX cases and gear into the mainstream: Games and Gamers! But going back to DIY, at least with computers, you don't need government approval to make your very own computer... when the same doesn't apply to some other bigger things... »

Sooo, the more powerful cores we have, the better performance we get at "crunching numbers", but the communications system might not keep up with the processing units, AND we require serious cooling systems. But, if we trade 10 state-of-the-fastest-art cores for some 100 Atom like cores, will get a low heat expenditure, and the communications system might get at the same level... although it has 10 times more data to get around to the cores. Which means that these new "ant colony" like clusters are going to need some serious work done, software and hardware wise! Ah, the future... I can see it now... dandelion like wireless processing units: small, light as a dandelion and heat up as much as a firefly! »

This "cloud business" sounds to be a low profile concept of HPC, just a neat way to govern various computers into running distributed applications. In other words, don't bother the client computer with heavy work, just distribute the work over the network to others with more computing/storage power. Which is great, if I want to use a PDA (or just a wrist watch) and access all of my movies, music... and spreadsheets! I say this, because if it really was "The Solution for HPC", it should have a seriously cool name, like: "hurricane enabled", "typhoon enabled" or "cyclone grid", or even "gridded thunderstorm(s)" (with overclocking enabled on demand). Personaly, I would prefer "Moon Enabled", with really low latency, low temperatures, a whole satellite for storage, sun powered, outer space cooled and... well, no O2 to react with, to create green house effects! As for the CherryPal, seems like the perfect green solution, including the capability of reducing software/media piracy to 0% in their boxes! »

So basically, this means that the "traffic jam shockwave" (youtube it) also happens in HPC... even if the acceleration is digital! So I can only deduce that the predictive and flock systems that are being developed for computer driven vehicles, will also be used in HPC in a not so far future! Or vice-versa even!! Uhm... I guess hierarchical management agents might get to be introduced, instead of a single master in the cluster... thus relieving the interconnects with alternating data transfers... Nonetheless, one thing that can help in situations of cores being faster than the interconnects, is to use "real-time" data compression... I hope this is an hypothesis that can be taken into account!! Since you've got CPU to spend... think compact! It's not the ideal solution, but saves a bundle in money while you can't get the new "hyper-speed" network, and the current one was just a "bad buy". But hey, what do I know, I'm just a newbie in HPC! »