Super Computers Talk

Power Consumption – How will it shape the future?

What will be the new breakthrough that allows supercomputers to reduce their electric bills?

In present times, the world's fastest supercomputers have upkeeps in the millions. The energy bill alone is enough to make me cringe. In a time when going bigger seems to be synonymous with going faster, what will be the ceiling for this power play?

The cookie cutter solution for joining the supercomputer top 500 chart seems to be some combination of thousands and thousands of processors, some large number of accelerators, and a big, cool place to hold them all. At a certain point, I imagine energy consumption will slow down the advancement of supercomputers. Energy consumption is probably already a concern. I have read that many large datacenters (Facebook, Google, yadda yaddda) already supplement their own power consumption with solar energy or wind energy (in addition to finding greener solutions to cooling). But you don't hear as much about similar solutions in supercomputing. It could be that the big research companies are a bit more secretive because of the competition. I don't know.

Either way, something is going to have to change. It's pretty common knowledge that power consumption is already affecting the basics of computing. Transistor count kept doubling every two years or eighteen months, as Moore predicted, until the power consumption on a single processor was too much to handle. Around that time, multiple core processors started showing up, each of which had lower power consumption, but by combining multiple cores the number of transistors on a single processor kept rising. So, what's next? What will be the new breakthrough that allows supercomputers to reduce their electric bills?

The CPU and GPU

When I read that researchers at North Carolina State University had achieved a 20% boost in performance for GPU/CPU setups, I was... well, a little underwhelmed. It seems like I have known for months that GPUs are used as floating-point accelerators in newer supercomputers. What seems to be difference about this new advancement comes down to location.

It may seem a little odd that moving two things closer together (even in computing) results in performance boosts. In the olden days (which is pretty much today, when it comes to technology) the GPU and CPU were either separate components, or at least worked independently. More recently, chip manufacturers are starting to include GPU and CPU cores on the same chip. To put it simply, these researchers took what the cores were best at (CPU for fetching and GPU for processing) and made them do just that.

I don't mean to belittle the work of these researchers; their work is impressive, and I could not do it, but something else bothers me about these results. In this article, it mentions that GPUs and CPUs fetch information at about the same rate, but GPUs process data faster. Why don't they just use GPUs?

I vaguely recall GPUs being used as floating-point accelerators, which may be the reason they can't replace CPUs outright. There are other data types that have to be taken care of, and maybe that's where the CPU retains its value. But if that's the case, why not let the “CPU” (not so central) fetch all the data, and have accelerators for each data type?

The Supercomputer Graveyard - Where will they all go?

For years, we normalcomputer owners have been hearing about supercomputers and the research they have aided. But with the passing of time and the always rising performance of processors, many supercomputers have been left in the dust. Where will they all go?

With many home computer, it is relatively easy to boost performance by replacing the CPU or adding a gig or two of RAM. But with the top computers in the world, the CPU count is quickly approaching 100k (not even including accelerators); they might as well replace the whole thing. And that's exactly what they do.

I'm not the first person to wonder what will happen to the old supercomputers; Frank Munger of the Knoxville News Sentinel wrote about ORNL's two soon to be outdated computers. They may split them up and give sections to universities, but nothing is certain. Which brings another question to mind: what could they do with them?

With the ever-rising importance of "the cloud," it is not totally unreasonable for them to be sold or donated to businesses selling cloud computing. Or they could be given to charities that need them (who doesn't need some free processors?). They could be combined with some other mildly outdated hardware and sent to low-income schools. Along the same lines, they could send smaller pieces to high schools to cultivate young computer scientist and mathematicians.

There have to be better ideas out there; I have only named a few off the top of my head. What do you think?