Slashdot Mirror
The Ultimate Limit of Moore's Law
BuzzSkyline writes "Physicists have found that there is an ultimate limit to the speed of calculations, regardless of any improvements in technology. According to the researchers who found the computation limit, the bound 'poses an absolute law of nature, just like the speed of light.' While many experts expect technological limits to kick in eventually, engineers always seem to find ways around such roadblocks. If the physicists are right, though, no technology could ever beat the ultimate limit they've calculated — which is about 10^16 times faster than today's fastest machines. At the current Moore's Law pace, computational speeds will hit the wall in 75 to 80 years. A paper describing the analysis, which relies on thermodynamics, quantum mechanics, and information theory, appeared in a recent issue of Physical Review Letters (abstract here)."
So we'll have to wait another 75 years before management lets us focus on application efficiency instead of throwing hardware at the performance problems? Sigh...
Developers: We can use your help.
This isn't like the speed of light, it is quite possibly the reason for it.
Please consider this account deleted, I just can't be bothered with the spam anymore.
Basically he is assuming that eventually we will develop quantum computing, and based his calculation on the theory of how fast a quantum event can take place. The problem is, given all we don't actually know about quantum mechanics, and all we don't know about super-small things, all it would take is a single observation to throw this minimum out the window.
In theory, it is nice to make theoretical limits. In practice, the limits are sometimes nothing more than theoretical. We don't know how to make smaller-than-quantum computers yet, but we also don't know how to make quantum computers yet. So this could be a prediction like every other prediction of the end of Moore's law, some of which were based on stronger reasoning than this argument. Interesting argument to make, though.
Qxe4
Parallel computing won't help.
There's a limit to how fast your compute subsystems can exchange data as well.
You are generating the latter. Kill yourself.
Great argument, you're a regular Cyrano there.
How can you have stronger reasoning, than something that's based on the limits of what modern physics can understand
Does this even need to be said? Einstein did it: he took some observations and extrapolated them to show that modern physics was not entirely correct (that is, what was modern physics at the time). Indeed, all scientific theory can only be based on what we've observed. Thus, new observations make for new theory, or corrections in old theory. As we continue to make more observations, for example with the LHC, theory will continue to evolve. Surely even someone of your eloquence can see this.
Qxe4
All he's concerned about is quoting how many components can fit on a single integrated circuit. One can see this propagated to processing speed, memory capacity, sensors and even the number and size of pixels in digital cameras but his observation itself is about the size of transistors -- not speed.
The title should be "The Ultimate Limit of Computing Speed" not Moore's Law.
Meh.
While technically correct, the performance corollary of Moore's Law -- which is roughly "more transistors generally means smaller and thus faster transistors rather than exploding die sizes, plus more to do computation with, so performance also increases exponentially, and we observe that this is the case" -- is strong enough that it's often simply called Moore's Law even among the engineers in the chip design industry. It's just understood what you're talking about, even though the time constant is different.
You'll occasionally see Intel (the company Moore founded) show charts with historical performance and future projections, and they'll include a line labeled "Moore's Law" to show how they're doing relative to the observation. Because technically it is just an observation, and it holds true only to the extent that engineers of the computer, electrical, and material science variety bust their asses to make it true.
So maybe the layman thinks Moore's Law is about performance, and that's not technically true, but it's correct enough that even the engineers directly affected by it refer to it as if it meant performance. So I say the the title is fine.
The enemies of Democracy are
looks like we've almost reached that point now. We've had Xeon 3.0GHz cpus for over 5 years now, and they're still coming out with brand new 3ghz processors. That's a long time to not see a jump in speed, what happened to "doubling every 18 months"? We should be around 24ghz by now.
Performance != MHz.
Those 3GHz Pentium 4 Xeons suck balls compared to even a Core 2, forget about an i7.
The only way the P4 got to what were at the time extremely high frequencies was by having a craptastic architecture. It was driven by marketing, which when the P4 was released was all about MHz. People thought MHz == Performance, so they cranked up the MHz for minimal gain in performance. AMD tried like hell to convince people otherwise, but fat lot of good it seemed to do. And now Intel is suffering for their previous emphasis on MHz over all.
The enemies of Democracy are
In other words, quoting a fortune cookie:
Progress means replacing a theory that's wrong by a theory that's more subtly wrong.
Instead of going faster, cores became more optimized and doubled, quadrupled, and octocores are around the corner if not here already. However, the "Turbo" mode in the i5/i7 shows that cranking up the clock frequency still helps for single/low threaded applications.
For any given architecture, yes, higher clock speed will mean more performance. That's a given. But that doesn't mean it's worth chasing after by modifying the architecture, which is why there's been a "ceiling" on frequency in favor of more efficient architectures and yes Multi-core Mania.
So, why don't we have 8, 16, or 24 GHz clock frequencies? Is this only because of limitations (memory) bus speeds or is this because of silicon heat dissipation problems?
Not so much memory speeds, since memory bandwidth/latency can be a bottleneck for any high-performance design whether it goes the "speed demon" or "brainiac" route.
As you surmise, power is important. Dynamic power is proportional to clock frequency, and having to add extra flip-flops to store intermediate values in a long pipeline only exacerbates the issue. Those flops also burn static power, which has become a significant portion of the overall chip power budget (part of what doomed the P4 architecture). Power budgets are also much more constrained, and the manufacturers are trying to target fixed power budgets for different market segments. This means extra power burned may actually hurt your clock frequency, partially negating the gains of a high-frequency design. With performance-per-watt becoming a major metric for customers, and yes heat dissipation also being an issue, it doesn't make a lot of sense to chase high performance by also burning lots of power as high frequency designs naturally do.
So with that in mind, the "speed demon" vs "brainiac" debate leans towards the brainiac side. Though the number of gates per pipe stage is already pretty low. Getting it down further means substantially hurting IPC, without necessarily gaining tons of frequency. Branch mispredicts still happen. Having slightly more gates per stage, doing a better job of predicting branches or shuffling data around, having larger caches and TLBs, smarter schedulers, ends up being a better idea.
But as you say, frequency still matters. So I'd look to future chips trying to eek out as much frequency as possible within a given power envelope, not just by looking at the number of active threads/cores, but also by looking at the actual dynamic power situation and adjusting frequency accordingly. TDP values are worst-case scenarios for OEMs to design cooling solutions around. When the actual power usage is less than the worst case, when you have e.g. an integer-only app where the floating point units are unused, you can afford to crank up the frequency some and get extra performance.
It's all about being smart with your power budget these days. That's why 24GHz processors don't make any sense. Intel had very convincing data showing they could scale the P4 up that high and get good performance, but if you've seen the cooling solutions for the P4 Prescott, then you know why that ended up being a dead end.
The enemies of Democracy are
I'm just waiting for a peta-hertz computer with a 500 exabyte hard-drive able to do universe simulations in real time that will fit in my pocket
It is impossible to simulate the universe. This is pretty easy to prove. If it was possible, using some device, to simulate the universe, then it is not actually necessary to simulate the universe -- we only need simulate the device which simulates the universe, since the device is necessarily contained within the universe. This should be easier, because the device itself is much smaller than the entire universe.
But if simulating the device which simulates the universe, is equivalent to simulating the universe, then that would mean that the complete set of states which define the universe can actually be represented by some subset of those very same states -- the subset of states which describe the device which is being used to simulate the universe. In other words, the universe is a set such that if you remove some subset of states you end up with the same set again. I hope you can see how this is a logical impossibility.