Here is my tweet on this subject
#AWS #SC10 rank 230 breaks a new era in #HPC It has billing on demand All #TOP500 entries must disclose $/yr/Rmax, or be extinct dinosaursThe point I wanted to make is in HPC, a price of $ 1.60 per node per hour is deceiving. To make an HPC configuration able to get on #230 on TOP500,we need so many nodes like never before, and pricing was not devised for these kind of volumes. The 25M assumes 100% usage, which is unrealistic and assumes on-demand prices, when one can use reserved instances pricing. This combined effect of lower utilization and the usage of reserves instances may reduce the cost to $6M to $8M, IMHO
But in Top 500, the annual operating cost are mind boggling. Here are the example of the cost of supercomputers (back in 2005)
As you see the combined cost of The Blue Gene/L (Rank 1) Supercomputer together with the ASC Purple (Rank 3) cost $290 million. If take a 5 year depreciation , that $58M per year. Add support Sysadmin and power and rent, at least 40% more, that is $81 per year to operate. Sure the Rmax for this computer is higher than Amazon, but Amazon @ say $8M per year in this context is 90% cheaper
If Amazon does something well, the best, is billing. As soon anyone will use AWS as a supercomputer,we will figure out very easily the exact total cost of an AWS supercomputer.
So AWS will do for Supercomputing, what it does now for the the Data Center, e which are transforming gradually in clouds. It forces private clouds to do metering and billing to justify themselves. Otherwise any DC risks to be des-intermediated and use hosting cloud companies A supercomputer is nothing but a huge private cloud, and now as AWS entered their territory, all this gigantic , dinosaur-size, TOP500 supercomputers will need to justify themselves economically. This is a real nightmare for the hard core in this industry, but it also will mean the democratic access to supercomputing applications we always wanted to achieve.
Amazon will soon improve the pricing to make CCI and Cluster GPU instances economically feasible on large scale. Knowing how amazingly they respond, I predict soon the special pricing will be in place to supercomputing even more affordable day by day.