The End of Supercomputers As We Know Them Today
About four years ago I predicted in this blog the fall of Grid Computing and the rise of Cloud Computing ( http://telzur.blogspot.com/2007/10/end-of-grid-computing.html )
Now I want to claim that Supercomputers as we know them today will not be the kind we will see in the future.
I think that within a couple of years:
1. Most supercomputers will shift to the cloud (HPC Cloud).
2. The governance of X86 is not guaranteed: We already see the rise of GPGPUs and we will see more families of technologies being involved, e.g. Tegra and ARM, Reconfigureable Computing (RC) i.e. FPGAs, etc'.
3. Supercomputers won't be anymore General Purpose machines. They will be reconstructed as a Cloud Service (Supercomputing as a Service) from all the available technologies (see previous section) per single mission (program) in order to maximize the performance and minimize the consumed energy and cost. I decided to use the word mission because the word task is usually reserved for a thread inside a program. Also the network topology will be ad-hoc (fully non-blocking, partially blocking, hyper-cube,....) and the user will be able to select from a few options the interconnect kind (Infiniband, 10GigE, ...).
Now I want to claim that Supercomputers as we know them today will not be the kind we will see in the future.
I think that within a couple of years:
1. Most supercomputers will shift to the cloud (HPC Cloud).
2. The governance of X86 is not guaranteed: We already see the rise of GPGPUs and we will see more families of technologies being involved, e.g. Tegra and ARM, Reconfigureable Computing (RC) i.e. FPGAs, etc'.
3. Supercomputers won't be anymore General Purpose machines. They will be reconstructed as a Cloud Service (Supercomputing as a Service) from all the available technologies (see previous section) per single mission (program) in order to maximize the performance and minimize the consumed energy and cost. I decided to use the word mission because the word task is usually reserved for a thread inside a program. Also the network topology will be ad-hoc (fully non-blocking, partially blocking, hyper-cube,....) and the user will be able to select from a few options the interconnect kind (Infiniband, 10GigE, ...).
Comments