Power-mad HPC fans told: No exascale for you - for at least 8 years
And here's why...
SaaS data loss: The problem you didn’t know you had
I recently stumbled upon a transcript from a very recent interview with HPC luminaries Jack Dongarra (University of Tennessee, Oak Ridge, Top500 list) and Horst Simon (deputy director at Lawrence Berkeley National Lab.) The topic? Nothing less than the future of supercomputers. These are pretty good guys to ask, since they’re both intimately involved with designing, building, and using some of the largest supercomputers to ever walk the earth.
The conversation, transcribed into a chat format in Science magazine, focused on the biggest challenge to supercomputing: power consumption. We can’t simply scale today’s petascale systems up into exascale territory – the electrical demands are just too much. The current top super, ORNL’s Titan, needs a little more than 8 megawatts to deliver almost 18 petaflops. If we scaled Titan’s tech to exascale (which means growing it by 18 times), we’d see power consumption at a whopping 144 megawatts – which, if you could even get it into the building, would cost something like $450m per year at current rates.
We’ve often heard that exascale systems will need to come in at 20 megawatts or less in order to be somewhat affordable. While the evolutionary improvements in power consumption have been significant over the last several years, they won’t be nearly enough to get us into that 20MW power envelope. In the interview, Dongarra and Simon talk about how we’re going to need some revolutionary technology (they mentioned stacked memory and optical interconnects) to get us to a point where we can even talk about firing up an exascale system.
In the words of NVIDIA chief executive Jen-Hsun Huang: "Power is now the limiter of every computing platform, from cellphones to PCs and even data centres." But power consumption is only the first, and highest profile, problem. They say we’re going to need to see changes – even breakthrough – on several fronts, including operating systems, applications, and even algorithms in order to bring exascale home. And breakthroughs aren’t free, nor even very cheap. Simon said that a “complete exascale program” could cost an additional $300m to $400m per year for 10 years – over and above what is being spent on HPC now.
Given the current economic climate, it isn’t surprising to learn that funding, at least from Western nations, isn’t hitting these levels. Which is why neither of these HPC authorities is betting on exascale by 2020.
There’s plenty more interesting discussion in the interview, including China’s changing role in HPC, the benefits of exascale and the way HPC technology trickles down into even consumer products. And with all that Really Big Data on the way, we could all soon be indirect beneficiaries of all the funds and research various companies and governments have invested in it. ®
COMMENTS
Just build a nuclear plant.
Of course research leading to more power efficient HPC needs to continue, but what's wrong with building a new 2GW power plant (a quick googling says ~$2-5 bn) with a new exa-number cruncher on the side. The extra power could be sold to help offset the computing hardware costs, and as HPC efficiency improves, more computing capacity could be added. Not to mention that the life time of the plant far exceeds that of any particular HPC setup, and it you get the added bonus of packing that many more smart people under one roof (physicists/mathematicians/comp engineers etc). To get around the NIMBY/eco-nut problem just tack the whole thing onto an existing research lab such as Oak Ridge or Sandia National Laboratories.
20W for that much neural computing horsepower...
To me, it seems pretty wasteful to emulate a rather fuzzy/analog neural network on a razor-sharp vaguely von-Neumannian architecture (albeit somewhat massively parallel). Perhaps just as wasteful as trying to teach a human brain (a vast neural network) to do razor-sharp formal logic - with all its massive ability to create fuzzy and biased associative memories and search/walk the associations afterwards, with all its "common sense" being often at odds with strictly logical reasoning ;-)
And for power mad HTPC fans, no exascale for you either because you're f****** bananas and keep dribbling into the fan vents and chewing on the cat.
IT infrastructure monitoring strategies
Agentless Backup is Not a Myth
Top 10 SIEM implementer’s checklist
Steps to Take Before Choosing a Business Continuity Partner
Enabling efficient data center monitoring
