Commoditization in HPC

Chris over at the excellent hpcanswers.com site wrote an article for HPCWire entitled “Innovation and Commoditization in HPC”. He makes quite a few points in there, but they have a constant theme.

The idea is to do things better than before, drive down your costs, increase the performance, and do this continuously. He sites several examples of this, and postulates others. This is a good article, though I would probably suggest refining some of the points. Specifically he points out that coprocessors are a good strategy. I am less convinced that co-processors are a good strategy in and of themselves, but are an element of a more broad based acceleration strategy; that is, it is hard to build a business on chips alone, but combine them with problems and solutions which have value, and you can do good things. Lots of folks out there build chips. Few marry them with apps of real value. The other issue is that in a commodity market, prices tend to fall (and occasionally rise with higher demand levels) over time. Producers find lower and lower cost methods to bring the system to market. Hence this suggests that as HPC has become largely commoditized in the cluster space, that adding decidedly non-commodity computing elements will be anathema to the market, and limit the overall market size/growth rate. HPC compute elements are largely disposable. $5k will buy you lots of processing power. If it dies, it dies, and you replace $5k of parts. This is curiously also why things like 3 year parts and labor in 4 hour response support contracts make almost no sense for such systems … it is far cheaper to purchase an additional unit in most cases, and have it sit on a shelf somewhere. Or in a rack, not plugged in. Add in virtualization, a good storage system, and if a node fails, restart it on a different element of your computing cloud, and swap at the bad hardware at your leisure. If your coprocessor will come in under the cost of a compute node, yet deliver 10x or better performance, this gets very interesting. I showed, at the Eilat conference, this little sub-node-cost hardware besting an 8 CPU cluster on a particular calculation. Neat stuff. Some of these things will work really well. Some will not. But they have to be cheap and easy enough to use.