Monday, May 13, 2013

No Exascale for You: Why We Will Not See an Exaflop System by 2020

The link to the presentation is in the title (like always). 

All of you who have been reading my blog for some time have noted that I am anti-Singularitarian.  I think the entire bit is bogus to the extreme.  I've talked about the 'Heat Death of the Singularity' and I have talked of the eye rolling Charlie's Accelerando (note: work is the distillation of the fantasies of the /. crowd) has induced (it stopped...eventually). I am also not concerned whatsoever about the Robopocalypse, my tag to the effect is as much a bit of humor as anything.

However, I'm a low level worker in the HPC field.  While, yes, I do move and shake a bit in the NewSpace world - its a much smaller one - I am not someone who, by career choice, is can have a big impact with my opinions in the world of supercomputing.  

Horst Simon, on the other hand, as a Deputy Director of Lawrence Berkeley National Lab, has far more impact when he stands up and says something.  He just has.  We will not be getting to exaflop machines by 2020, he says.  That takes guts to say.  Go read his prezo on the subject.  I wish there was a video of him presenting a version of it at the CUG LBNL just hosted.  I've been told it caused quite a stir.


Josh Cogliati said...

Why would anything in the presentation preclude the singularity? (I only saw the slides, not the talk.)

The statement "Even under best assumptions in 2020 our brain will still be a million times more power efficient" is a statement about how much power it takes to simulate a human brain, not how much power it takes for a computer to have general intelligence that is greater than a human. I would be incapable of simulating the Z3 computer at realtime (with only my brain, paper and pencil) let alone a modern computer, but that does not mean that computers were smarter than humans back in 1941.

Will Baird said...

All tech goes through a sigmoid. Projected in here is - barring a major breakthrough - we are nearly at the top of the S.

I'll write up a detailed discussion of it in the future.

Josh Cogliati said...

Well, if nature can make a brain run on 20 W than humans should be able to achieve the same thing. The human brain is an existence proof of what is possible.

Admittedly the brain and computers work very differently. The brain tends to rewire things (hook up synapses) to do long term memory and programming, where as computers (I'll ignore FPGA's for the moment) use data for the programming and memory (a flip flop or a DRAM cell continually use power to store data). Programming brain-like hardware would be very different than programming a von Neumann architecture computer.

Josh Cogliati said...

Oh, and I look forward to a detailed discussion of how computer technology is near the top of the sigmoid.