HPC Wire states that Zeta flops will occur says some analyst
While Thomas Sterling’s interview about the impossibility of reaching zettaflops made a lot of sense, the history of making negative predictions about technology is often an embarrassing one. Here are three examples: “I think…
Those papers are too generic ; I would like to see mathematical logic to explain why it is possible or not ;; for exemple what is the limit of optics? 1fJ/bit?
It is clear there is a physical limit to computation, this is very sad, but we are still very far from it, there are plenty potential solutions to extend the Moore Law
Is it an econimical limit? power limit? A communication limit? Complexity limit? For what problems? a machine to treat 1 problem or thousands/millions in parallel?
There is a clear potential for sciences to use a reasonably cheap zettaflop machine
It can be 10B$ “this is very cheap” regarding our deficits, it can be 1GW, just 1 nuclear plant ; it all depend what fundamental problems this machine can solve for humanity that we cannot be solved without it
I think the true limit is the ecomical balance between cost/benefits , we will reach that limit before the true physical limits
Meanwhile we should not neglect those fundamental limits that are on our radar, I feel that in future we will be more limited by data storing and transport capability than compute capability mainly because the data will come from of trillions of sensors dispersed on our planet
A very interesting paper that does discuss some of the various physical contstraints on a zettaflop machine can be found here: http://www.cse.nd.edu/Reports/2006/TR-2006-15.pdf It is from 2006, would love to see an update on where we have moved along the path (or not)
For me it is very clear, that we will reach Zeta OPERATIONS computers. The difference I see is, that the demand WHAT calculations to perform will change as technology enhances. So it could be that there is no demand for floating point numbers anymore. maybe we want to simulate a brain or so and simply store and calculate analogous relative information. So the elements of such computer need not to be electrical or light pulses but maybe chemical molecules.
Thoughts like this are a bit philosophical but the development of technology is exponential and thus the thoughts should consider this.
When the question is
“Are we able to build a Zettaflops computer using traditional components which are available in 20 years ?”
even then I feel confident that we (mankind) could do it. Structures of about 5nm (4 shrinks from now on, 10years?) could help to reduce power demands as well to increase processing capabilities. From 28nm to 5nm we will have an increase of round about 32 in area. Power demand maybe one fourth of today’s (?)==>
Today we are able to build up chips which contain 10k FPUs per Die in about 100mm² without control logic/caches around (maybe more, but it’s a rough estimation). So let’s double it ==> 200mm².
This thing runs @ 2Ghz and uses ~200W power. ==> 10.000 FPUs * 2Ghz => 20TF/Die/200W.
1.0 ZettaFLOPs (10^21 FLOP) =
Today: 50M Dies (today), 10GW Power
in 10 years: ~1M Dies, 2.5GW Power
in 20 years: ~50k Dies, 10MW Power (200W/Chip)
!!!VERY ROUGH estimation!!!
My brain starts boiling when I think about zetaflops. I’d better wait for exaflops first, just to stop shivering when I try to predict all aspects 🙂
For people that need to make huge investments in profund code transformations they need to understand what is behind exascale ; where is the limit? are the transformations viables over 20 years? or 5?
If the limit is exaflop or just behind let’s say 10EF ; then we need to do a huge investment to focus on software optimisations etc
If the limit is 1EF then we have a little more time; but clearly one day or another we will have to focus on maths and algorithms
Exaflop will be reached : not doubt between 2016 or 2022 ; will depend on political and economical environment
We can expect some country will make a Exaflop for linpack and be 1st on top500,
it will be very expensive, and very inefficient
Then 2 to 5 years after we will see an industrial exaflop machine ; it will be radically different “maybe surprising” and focused on high throughput computing, very economical, efficient and useful for our society and industries
NOTE I now post my TRADING ALERTS into my personal FACEBOOK ACCOUNT and TWITTER. Don't worry as I don't post stupid cat videos or what I eat!