Can Engineers Overcome the Impending Limitations of Moore's Law?

[ElliotSmith]

The scientific limit on to how small you can make a functionally viable transistor is very fast approaching and should hit a stone wall within the next 10 years or less. How will electronic engineers and computer scientists compensate for this problem?

Without some revolutionary breakthrough with the design of and fundamental basics of microprocessor chips, I don't see how they can do this. Are there any workarounds on the table being discussed and researched for this issue?

This is why commercially-available CPU's from Intel and AMD have gone in the direction of energy efficiency instead of sheer performance. It's becoming very difficult to squeeze more and more performance out of each generation of microarchitecture without improvising like tweaking instruction sets, adding more cache, and making memory bandwidth and motherboard chipset improvements.

 

[rcgldr]

Unless there's some breakthrough in the future, once some type of physical limit is reached, there's not much that can be done. I'm not sure if the limit is due to transistor size itself or in the process of producing a chip. It seems that current methods include having more layers on a chip, but then there needs to be a way to dissipate the heat for the inner layers.

Another issue is that in order to increase speed, you need a higher voltage to gate size ratio, but this creates a localized heat problem that is easiest to solve by reducing gate density (more space between gates to dissipate heat), but lower density means larger chips and/or reduced gate count which translates into increased cost. This is why consumer oriented processors have been stuck at around 4 ghz unless liquid cooling is used.

 

[ElliotSmith]

Do you think it's possible that quantum computers will be available for retail purchase and be commonplace in most households at some point probably in the distant futrue?

Also, is it possible to make a room-temperature quantum computer? Currently, these computers have to be cooled to a temperature almost absolute zero with liquid helium in order to maintain the quantum coherence of the qubits. This would be highly impractical for household use.

However, topological insulators in quantum processors would solve this issue in theory. But actually creating a topological quantum computer has proven to be a daunting challenge for scientists and engineers.

Quantum computing is still in it's infancy and won't be perfected for quite some time.

 

[ricardo81]

>Quantum computing is still in it's infancy and won't be perfected for quite some time.

Seems to be the sum of it for now, apparently there are mounting examples of quantum effects in nature so it does indeed seem possibly to perform quantum computation at a much higher temperature. I believe the efficiency of photosynthesis is put down to quantum effects.

 

[jack476]

Well, there are certainly things you can still do at the level of the CPU's architecture or the implementation before we start getting into technology that is still highly speculative and experimental and won't be available to the end user for a long time, if ever. The individual components on the IC can only get so small, but that doesn't mean you can't just keep adding more of them to a more efficient architecture, or have better cooling and more efficient software.

 

[jim mcnamara]

Commodity consumer cpu chips - what the discussion is about - are benefitting from improved compilers. Intel (x86) and Fujitsu (sparc) have special opcodes and specialty libraries to take full advantage of a given cpu architecture. Since gcc (Linux/Android/Consumer electronics) has made strides in this arena, Microsoft is very likely doing just that - improving Visual Studio compilers.

 

[Mark44]

As I understand things, the "wall" is not 10 years out - it has already happened. The CPUs haven't increased in clock speed for the past several years, and seem to be running at a maximum of about 3 - 3.5 GHz. Feature size inside a chip is closely tied to how fast the chip can run, as the smaller the individual components inside a chip, the closer together they are, and the faster information can be transferred.

To compensate for this inability to produce chips at finer resolutions, manufacturers such as Intel and AMD have put putting more CPUs in a single chip. My current desktop, which I bought about a year ago, is running a Quad Core Duo processor, which has eight virtual CPUs.

 

[timthereaper]

The problem with transcending Moore's law is that there isn't really a very pressing need yet. Sure, some industries need a whole whopping amount of computing power, but I believe necessity really is the mother of invention. When things really started to take off in the 1970s, there were a lot of strides that needed to be made to get computers into the hands of the everyday person or to make them good enough to complete a necessary task. People came up with inventive techniques for chip manufacturing and design because there was a need for a faster machine or a more powerful one. The demands of the common user drove software companies to include more features, increasing program and OS size and requiring better machines.

I would make the argument that the user base of the world at large is satisfied with the speed and capabilities of the modern computer. Most of the development in computing currently is in creating better, leaner, and more intelligent software that takes advantage of the available hardware. In contrast to even 20 years ago, the common user doesn't have to worry about running out of memory or disk space too often. The speed issue today is about how fast we expect things should run, not that things are really too slow (more an issue of impatience than necessity). However, like I said before, there are many companies out there trying to solve problems of immense complexity and require more computing power, but I would argue they are the only ones really driving the increases in hardware tech.

We'll only get around our limitations when transistors stop being the base technology of computing. It's really only been 75 years since Shannon published his thesis linking Boolean algebra to computing. The transistor was only invented 65 years ago. We've been basically iterating on that train of thought since then and we may have already reached the end of the tracks. There might be other directions computing can take that we haven't ever thought of yet. Quantum computing is a great advance in a slightly different direction and I hope it takes off. However, I think that only when we really need the computing power will we see that kind of computing growth rate again.

 

[nsaspook]

As I understand things, the "wall" is not 10 years out - it has already happened. The CPUs haven't increased in clock speed for the past several years, and seem to be running at a maximum of about 3 - 3.5 GHz. Feature size inside a chip is closely tied to how fast the chip can run, as the smaller the individual components inside a chip, the closer together they are, and the faster information can be transferred.

The 'wall' is still out there with what can be done even with old man silicon. The Intels of the world have to make money off the massive investments in current technology.
http://semiengineering.com/will-7nm-and-5nm-really-happen/

 

[Mark44]

Also, I was reading up on the next generation of Intel CPUs - Haswell architecture. I don't recall the feature size, but the clock rate was around 4.6 Ghz or so.

I remember reading some time back that a big concern with the smaller feature size was quantum effects due to electrons "tunneling" to different levels. I haven't seen anything about that lately, but then again, haven't been following that closely.

 

[ElliotSmith]

I read in Maximum PC magazine that AMD plans on releasing an air-cooled quad core chip with a default clock speed of 5GHz.

Not sure how they can get away with air cooling @5GHz, that's usually the cutoff for requiring liquid nitrogen.

AMD's flagship CPU is usually only about half as fast as Intel's.

 

[TheDemx27]

IMHO I don't think quantum computing will really be pushing moore's law. From what I've heard quantum computers aren't really that great for applications we are used to and are only good for doing certain operations i.e. finding prime factors. I can't see why the quantum computer would become more than a commercial product. There isn't any reason to put a quantum computer in the hands of the average Joe, as he won't have many uses for it.

 

[phinds]

... There isn't any reason to put a quantum computer in the hands of the average Joe, as he won't have many uses for it.

Right. And Bill Gates was SURE that 64K would be all the memory anyone would ever need. It was just inconceivable that more could be required for a single person.

 

[SteamKing]

IMHO I don't think quantum computing will really be pushing moore's law. From what I've heard quantum computers aren't really that great for applications we are used to and are only good for doing certain operations i.e. finding prime factors. I can't see why the quantum computer would become more than a commercial product. There isn't any reason to put a quantum computer in the hands of the average Joe, as he won't have many uses for it.

At one time, the chairman of IBM (!) predicted that the market for general purpose computers was too limited to warrant investment by the company into their production, or so the story goes. Large mainframe computers quickly evolved into more affordable mini-computers, thence to super minis, work stations, and finally desktops, portables, laptops, and tablets, all within one lifetime. Sure, a lot of the things we use computers for are completely mundane, but that doesn't mean that society hasn't changed as a result.

Right now, the applications of quantum computing may seem non-existent, but once they are built, who knows? They could lead to a Hacker Apocalypse, where no conventional system is safe from being plundered.

 

[SixNein]

The scientific limit on to how small you can make a functionally viable transistor is very fast approaching and should hit a stone wall within the next 10 years or less. How will electronic engineers and computer scientists compensate for this problem?

Without some revolutionary breakthrough with the design of and fundamental basics of microprocessor chips, I don't see how they can do this. Are there any workarounds on the table being discussed and researched for this issue?

This is why commercially-available CPU's from Intel and AMD have gone in the direction of energy efficiency instead of sheer performance. It's becoming very difficult to squeeze more and more performance out of each generation of microarchitecture without improvising like tweaking instruction sets, adding more cache, and making memory bandwidth and motherboard chipset improvements.

The problem of course is quantum tunneling.

http://en.wikipedia.org/wiki/Quantum_tunnelling

At any rate, we will adjust by making things more efficient programming wise on current architectures.

The next big speedup won't come from changes to CPU frequencies; instead, it will come from new architectures. Architectures based on the brain for example could be very powerful.

The current chips are much faster than the human brain, but the human brain rapes chips on parallel processing. And it does it on very little energy.

 

[.Scott]

IMHO I don't think quantum computing will really be pushing Moore's law. From what I've heard quantum computers aren't really that great for applications we are used to and are only good for doing certain operations i.e. finding prime factors. I can't see why the quantum computer would become more than a commercial product. There isn't any reason to put a quantum computer in the hands of the average Joe, as he won't have many uses for it.

A quantum processor would be like a graphics processor. It would be tasked by a regular computer to solve problems it was exceptionally good at. There is no reason to think that the applications for a $2000 QM processor would be exclusively commercial.

 

[.Scott]

The problem with transcending Moore's law is that there isn't really a very pressing need yet.

I have been a software engineer for over four decades - before the term was invented. There has never been the perception of a "pressing need". Never-the-less, additional processing power and memory capacity has always been exploited.

 

[ElliotSmith]

What about topological quantum computing? As in room-temperature quantum computers without having to cool them to absolute zero degrees in order to maintain the coherence of the qubits.

Is it currently possible to make topological insulators?

 

[.Scott]

What about topological quantum computing?

QM information processing allows certain operations to be performed much faster than conventional computers. However, they are not general purpose in the same way that conventional computers are.

QM information processing will not represent a extension of time for Moore's Law. it will represent an entirely new direction in data processing.

As for your technology suggestions (ex, topological insulators), I don't doubt that some technology will be found.

 

[timthereaper]

I have been a software engineer for over four decades - before the term was invented. There has never been the perception of a "pressing need". Never-the-less, additional processing power and memory capacity has always been exploited.

Okay, so "pressing need" was a probably a bad choice of words. The computer wasn't built for anyone specific purpose, so there were no "needs". I concede that point.

What I meant was that we found more and more uses for computers but computing power has been, up until recently, very limiting. Gaming and print graphics, computer-aided design and engineering tools, computer animation, and other such applications were limited by the lack of computing power available. The gains from Moore's Law were more apparent with each new chipset that came out because there were more applications that could take advantage of the increase in speed. There was more incentive to be creative with chip design and new manufacturing processes. The hardware was more clearly the limitation of the technology, not the software. Contrast that with today. I'm not sure that the common user even notices the speed difference between their old machine and new one. As well, there's a general shift away from personal computers toward laptops, tablets, and smartphones (i.e. less powerful but more mobile hardware). I would argue processor speed for the common user is more now a "convenience" than a "necessity", like a weapon to combat software bloat. We strive for faster machines, but for what? Server farms and virtualization, problems involving Big Data, gaming, and maybe a few other things I'm sure I'm missing, but nothing that the average user concerns themselves with.

I would contend that unless we get something that the common user will need a heavier-duty processor for, we won't see the kinds of technological leaps that we got during the 1970s-1990s.

 

[SixNein]

Okay, so "pressing need" was a probably a bad choice of words. The computer wasn't built for anyone specific purpose, so there were no "needs". I concede that point.

At one time, it was built for arithmetic.

 

[RobS232]

The scientific limit on to how small you can make a functionally viable transistor is very fast approaching and should hit a stone wall within the next 10 years or less.

The prediction you have made has been made continuously for the last 20 years, yet engineers have discovered work-arounds for the optical limits that were thought by scientists to limit lithography resolution.

Gordon Moore never made a prediction on speed. His prediction referred to the number of transistors. Higher speed has been a by-product of decreasing transistor size. The number of transistors directly impacts the complexity of the functions that an integrated circuit can provide. Since electronic systems today can be large enough to utilize thousands of integrated circuits, there is no reason to believe that even higher transistor counts cannot be useful. Today ICs are being manufactured with billions of transistors.

 

[RobS232]

Unless there's some breakthrough in the future, once some type of physical limit is reached, there's not much that can be done. I'm not sure if the limit is due to transistor size itself or in the process of producing a chip. It seems that current methods include having more layers on a chip, but then there needs to be a way to dissipate the heat for the inner layers.

Another issue is that in order to increase speed, you need a higher voltage to gate size ratio, but this creates a localized heat problem that is easiest to solve by reducing gate density (more space between gates to dissipate heat), but lower density means larger chips and/or reduced gate count which translates into increased cost. This is why consumer oriented processors have been stuck at around 4 ghz unless liquid cooling is used.

Speed is increased by using shorter channels, thus increasing the field for a given voltage. Voltages have also bee reduced, trading off speed and power.

Silicon is an excellent thermal conductor. The critical path for heat transfer is the interface between the IC surface an the heat removal structure (often a heat pipe). Large area chips with distributed power consumption (i.e., multiprocessor chips) are very effective in dissipating heat.

 

[RobS232]

"I would contend that unless we get something that the common user will need a heavier-duty processor for, we won't see the kinds of technological leaps that we got during the 1970s-1990s."

Meanwhile, software engineers are simultaneously utilizing thousands of computers, sitting on farms at Google, Amazon... to carry out complex data analysis activities. A friend described his activity to me, which was utilizing 27,000 computers to process data in real time.

 

[timthereaper]

"I would contend that unless we get something that the common user will need a heavier-duty processor for, we won't see the kinds of technological leaps that we got during the 1970s-1990s."

Meanwhile, software engineers are simultaneously utilizing thousands of computers, sitting on farms at Google, Amazon... to carry out complex data analysis activities. A friend described his activity to me, which was utilizing 27,000 computers to process data in real time.

Keywords in that quote: "common user". I never said that no one needs a more powerful processor or that there's not a reason to develop one. In fact, I think I mentioned that those companies are the ones pushing the tech forward. However, the incentive isn't the same as during the 1970s-1990s. Although you can make a lot of money from large companies buying your products, you can get way more if you can get consumers to buy your product. Back in the day, practically everyone shelled out $1000+ every 2 years to upgrade to a better machine because you could count on it being measurably better and more capable than the old machine. Now, I'll bet the common user doesn't even notice the difference when they buy a new one.

 

[rcgldr]

One change is that there is a much greater number (and percentage of the population) of common users buying computes versus decades ago, and the majority of those buy computers or laptops costing much less then $1000 (USA).

 

[harborsparrow]

Decreasing size and stagnant speeds for processors is not new, but speeds are now great enough and sizes small enough that it is not a problem for many things. It may be that the emphasis in computer architecture has shifted to power savings, rather than more speed or less size.

 

[RobS232]

Keywords in that quote: "common user". I never said that no one needs a more powerful processor or that there's not a reason to develop one. In fact, I think I mentioned that those companies are the ones pushing the tech forward. However, the incentive isn't the same as during the 1970s-1990s. Although you can make a lot of money from large companies buying your products, you can get way more if you can get consumers to buy your product. Back in the day, practically everyone shelled out $1000+ every 2 years to upgrade to a better machine because you could count on it being measurably better and more capable than the old machine. Now, I'll bet the common user doesn't even notice the difference when they buy a new one.

First. with cloud computing, a larger fraction of the total number of processors sold go to big companies. The rest of us use them indirectly - for example, we are told what products we want to buy when we browse a web page .

However, there is an unmet need for significantly larger processing power than is currently available with a main stream consumer market - high resolution video image processing. Today's rapidly expanding capability to deliver video entertainment by streaming is limited by available bandwidth. This limitation can be dealt with by video compression, but quality suffers. More sophisticated compression algorithms can deliver increasingly high resolution video (HD today, 4K tomorrow,...) but they require real time processing of increasing magnitudes in power to operate in real time. Faster processors than we have today will be needed to meet this need. Faster does not mean clock speed (which has nothing to do with Moore's law, anyway - at least as defined by Moore). It has to do with MIPS or FLOPS, For a given architecture they are correlated, but can change dramatically with architecture changes. The accepted solution is to use Moore's law as he defined it - add more transistors. Large graphics processors use increasing numbers of parallel CPUs, all on the same chip. Video data is especially suitable for parallel processing. Your (or, at least my) TV set will be the market for more sophisticated processors.

Another market that will expand the use of high performance processors is artificial intelligence. For example, increasingly sophisticated systems will be needed to collect a wide range of environmental visual, motion and location data and make decisions for an intelligent auto.

 

[SixNein]

Decreasing size and stagnant speeds for processors is not new, but speeds are now great enough and sizes small enough that it is not a problem for many things. It may be that the emphasis in computer architecture has shifted to power savings, rather than more speed or less size.

The problem has bee in quantum tunneling.

 

[lpetrich]

I think that we are close to the limit for online video, if we have not already reached it. It's hard to justify much more than a few thousand pixels in each dimension.

But there are other applications that are likely to demand more and more CPU time.

3D simulation and graphics, especially in real time. This is important for video games, which have become a major form of entertainment. That's what's driven 3D rendering capability so far, and that's what's likely to continue to drive it in the near future. In addition to improved surface-reflectance modeling, one can expect improved lighting and improved entity physics. Doing fluids correctly will be a big challenge, however, since one needs to do simulations over a 3D grid or by turning them into a large number of blobs and then doing Smoothed Particle Hydrodynamics on them.

Artificial hearing. Perception of speech is the most important application here. It is something that we easily do in real time for languages that we know, but for computers, it is difficult. One can get a computer to recognize a few words that were spoken by many speakers or many words that were spoken by a few speakers, but it's difficult to get a computer to recognize many words spoken by many speakers. Part of the solution to this problem may simply be more computer cycles, to improve the pattern recognition.

Artificial vision. That has lots of uses, like in robotics and in self-driving cars.Looking at these applications, they typically involve a few operations repeated numerous times over many data sets. So SIMD CPU's are the most suitable for them, with general-purpose CPU's for overall control. SIMD is Single Instruction Multiple Data, something that simplifies their design. However, it is at the price of making control flow rather kludgy, like following both directions of a branch and then using the results of only one of them. SIMD is already widely used in video cards, so we can expect to see more of it.

 

[joema]

The scientific limit on to how small you can make a functionally viable transistor is very fast approaching and should hit a stone wall within the next 10 years or less. How will electronic engineers and computer scientists compensate for this problem?...Are there any workarounds on the table being discussed and researched for this issue?...

There are four main limits re CPU performance:

(1) Clock speed scaling, related to Dennard Scaling, which plateaued in the mid-2000s. This prevents much faster clock speeds: https://en.wikipedia.org/wiki/Dennard_scaling

(2) Hardware ILP (Instruction Level Parallelism) limits: A superscalar out-of-order CPU cannot execute more than approx. eight instructions in parallel. The latest CPUs (Haswell, IBM Power8) are already at this limit. You cannot go beyond about an 8-wide CPU because of several issues: dependency checking, register renaming, etc. These tasks escalate (at least) quadratically, and there's no way around them for a conventional out-of-order superscalar machine. There will likely never be a 16-wide superscalar out-of-order CPU.

(3) Software ILP limits on existing code: Even given infinite superscalar resources, existing code will typically not have over 8 independent instructions in any group. If the intrinsic parallelism isn't present in a single-threaded code path, nothing can be done. Newly-written software and compilers can theoretically generate higher ILP code but if the hardware is limited to 8, there's no compelling reason to undertake this.

(4) Multicore CPUs limited by (a) Heat: The highest-end Intel Xeon E5-2699 v3 has 18 cores but the clock speed of each core is limited by TDP: https://en.wikipedia.org/wiki/Thermal_design_power
(b) Amdahl's Law. As core counts increase to 18 and beyond, even a tiny fraction of serialized code will "poison" the speedup and cap improvement: https://en.wikipedia.org/wiki/Amdahl's_law
(c) Coding practices: It's harder to write effective multi-threaded code, however newer software frameworks help some.

While transistor scaling will continue for a while, increasingly heat will limit how much of that functional capacity can be simultaneously used. This is called the "dark silicon" problem. You can have lots of on-chip functionality but it cannot all be simultaneously be used. See paper "Dark Silicon and the end of Multicore Scaling": https://www.google.com/url?sa=t&rct...=k_D1De2gUp79VwMVcTIdwQ&bvm=bv.84349003,d.eXY

What can be done? There are several possibilities along different lines:

(1) Increasingly harness high transistor counts for specialized functional units. E.g, Intel core CPUs since Sandy Bridge have had a Quick Sync dedicated video transcoder: https://en.wikipedia.org/wiki/Intel_Quick_Sync_Video This is about 4-5x faster than other methods. Intel's Skylake CPU will have a greatly improved Quick Sync which handles many additional codecs. Given sufficient transistor budgets you can envision similar specialized units for diverse tasks. These could simply sit idle until called on, then render great performance in that narrow area. This general direction is integrated heterogeneous processing.

(2) Enhance existing instruction set with specialized instructions for justifiable cases. E.g, Intel Haswell CPUs have 256-bit vector instructions and Skylake will have AVX-512 instructions. In part due to these instructions a Xeon E5-2699 v3 can do about 800 linpack gigaflops, which is about 10,000 faster than the original Cray-1. Obviously that requires vectorization of code, but that's a well-established practice.

(3) Use more aggressive architectural methods to squeeze out additional single-thread performance. Although most items have already been exploited, a few are left, such as data speculation. Data speculation differs from control speculation, which is currently used to predict a branch. In theory data speculation could provide an additional 2x performance on single-threaded code, but it would require significant added complexity. See "Limits of Instruction Level Parallelism with Data Speculation": http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.47.9196&rep=rep1&type=pdf

(4) Use VLIW (Very Long Instruction Word) methods. This side steps the hardware limits on dependency checking, etc by doing it at compile time. In theory a progressively wider CPU could be designed as technology improves which could run single-threaded code 32 or more instructions wide. This approach was unsuccesfully attempted by Intel with Itanium and CPU architects still debate whether a fresh approach would work. A group at Stanford is actively pursuing bringing a VLIW-like CPU to the commercial market. It is called the Mill CPU: http://millcomputing.com/ VLIW approaches require software be re-written, but using conventional techniques and languages, not different paradigms like vectorization, multiple threads, etc.

 

[Carno Raar]

Right. And Bill Gates was SURE that 64K would be all the memory anyone would ever need. It was just inconceivable that more could be required for a single person.

He actually never said that but it's a popular urban legend that he did.

Back on topic, Moore's Law seems to be reaching the end of its life now. We're moving to distributed systems and multicore machines and Amdahl's Law is the new one to watch.

http://en.wikiquote.org/wiki/Bill_Gates

http://en.wikipedia.org/wiki/Amdahl's_law

 

[phinds]

He actually never said that but it's a popular urban legend that he did.

Either way my point remains exactly the same

 

[SixNein]

Amdahl's law is algorithm dependent. So it's not the same kind of thing as Moore's Law.

 

[Carno Raar]

Amdahl's law is algorithm dependent. So it's not the same kind of thing as Moore's Law.

It's an appropriate answer for the OP's question.

"The scientific limit on to how small you can make a functionally viable transistor is very fast approaching and should hit a stone wall within the next 10 years or less. How will electronic engineers and computer scientists compensate for this problem?"

A valid answer is we spin up more cloud instances and learn to write concurrent code. Right now Amdahl and Moore are limiting factors in the growth of large computer systems. Moore will doubtless become less important in the near future, while we're only just starting to get our heads around concurrency issues. I say concurrency not parallelism as I don't yet have access to properly parallel hardware ... :-)

 

[phinds]

I say concurrency not parallelism as I don't yet have access to properly parallel hardware ... :)

If you don't have parallel hardware, concurrency is just sequential but with extra overhead. That is, if you have a single-thread process in a single CPU and you make it multi-threaded but still on the single CPU, all you have done is add thread overhead.

 

[Carno Raar]

If you don't have parallel hardware, concurrency is just sequential but with extra overhead. That is, if you have a single-thread process in a single CPU and you make it multi-threaded but still on the single CPU, all you have done is add thread overhead.

Managing multiple downloads + many other use-cases.

"Take this list of URLs and download them all". You don't want to sit there doing nothing while your 1st and only download times out.

Edit: Yes you can implement this single threaded with async and a non-blocking downloader but that's a bit weird - and most libraries implement nonblocking download with threads anyway.

 

[phinds]

Managing multiple downloads + many other use-cases.

Good point. Thanks. I had not thought about I/O bound processes.