DOE: Federal spending necessary for exascale supercomputer development
Building a viable exascale supercomputer by 2022 will likely require at least $1 billion to $1.4 billion in funding and it can't be built unless federal agencies spend money on its development, says a June Energy Department report to Congress.
FierceGovernmentIT obtained an unredacted copy of the report (.pdf) through a Freedom of Information Act request. A practical exascale computer--one capable of 1018 operations per second and 1018 bytes of storage--will require advances in supercomputing greater than the improvements behind the current state-of-the-art petascale machines, which can execute operations on the order of at least 1015 per second.
The Energy Department achieved petascale by networking clusters of commodity processors and memory, the report says, an approach that if applied today to achieving an exaflop would consume more than a gigawatt of power, "roughly half the output of Hoover Dam," the report says.
An exaflop computer built a decade from now with capacity obtained through the normal pace of technological improvement and without the benefit of commercially risky investments would still require more than 200 megawatts of power at an estimated cost of $200 million to $300 million annually, it adds. It also would "have an extremely high failure rate and limited memory per processor, and would be difficult to program."
Exascale suffers the problem of memory density improving at a slower rate than processor logic. Supercomputers currently require 10 times more energy to bring two numbers from memory into the processor than to perform the subsequent operation itself. Energy estimates that by 2020, that ratio could reach 50 times more.
Moving to exascale will require addressing memory and data movement issues, Energy says, stating that it also needs to foster a feedback loop with outside researchers and the private sector to inform them of application needs but ensuring that application developers are aware of proposed architectures and operating changes.
Another technological problem will be how to cope with run time errors in a exascale machine. Since clock frequency is unlikely to change in the next decade, an exascale supercomputer will require more processing elements--about a billion in all. That'll have the effect of potentially increasing the frequency of errors and making their correction much more difficult.
Energy says it will divide responsibilities for exascale development between the National Nuclear Security Administration--a semi-autonomous agency within DOE--and the Office of Science. The office will focus on long lead-time research and development, while NNSA will have primary responsibility for systems engineering. An Exascale Coordination Board headed by an NNSA official with an Office of Science deputy will synchronize efforts and ensure accountability, the report says.
It adds that responses to a July 2011 request for information issued by Argonne National Laboratory at the department's request produced responses that consistently estimated the cost of developing an exascale system at between $1 billion and $1.4 billion.
At this time, however, "due to current and likely ongoing budget constraints and an evolving understanding of industry status and agency needs," Energy says it hasn't identified technical milestones nor their associated costs for achieving an exascale supercomputer.
- download the report, "Department of Energy Exascale Strategy, Report to Congress" (.pdf)