Another Step Toward the End of Moore’s Law | Hacker News Reader

Another Step Toward the End of Moore’s Law

spectrum.ieee.org

148 pointsfurcyd130 comments

Moore's law itself may be dead but there is still additional potential to explore. Right now, CPUs are mostly two dimensional with only a few layers stacked on top compared to the extent in the other two dimensions.

Compare this to a human brain which too, has a sulcated layering system for the grey matter but the white matter connectome turns it into a fully three dimensional object. We can't produce such a thing with technological means yet.

The CPUs we produce have to be deployed in datacenters and computers within distance to each other, mostly for heat dissipation reasons. We can't just stack $N CPUs onto and next to each other until we reach an object of the size of a brain. It would not be practical to cool that thing. Compare this to human brains which require much less energy (20 Watts on average).

Moore's law is dead but it's about size of transistors. Human axons have diameters of hundreds of nanometers, even the gap of intracellular space of a synapse is approximately 20 nanometers. Compare this to the 5 nanometers we got now. We don't need size improvements to model brains, we need power consumption/heat dissipation/electric loss improvements. Also, CPUs are still very expensive. We also need improvements in manufacturing them more cheaply. Both isn't really requiring node sizes to shrink further.

Sadly, there's a different law called Dennard's Scaling which was precisely about power consumption, but it has been dead for about 13 years. https://en.wikipedia.org/wiki/Dennard_scaling

maybe in-silicon micro-vias for heat exchange ?

The brain architecture is not the end-all of computing. For example weather prediction is just about fast math, a brain like architecture is useless here.

The main issue is that the death of moore's law somehow implies that there won't be any progress in chip manufacturing and we'll always use the computing technology we have now because at some point you reach sizes of single atoms. But this is wrong as we don't need node shrinks, we need different improvements. And human brains are the proof that they are physically possible.

Human brains, if you ignore brains of mammals larger than us, are the most complex "computers" we know of. And they are much more efficient at that and their assembly doesn't require huge buildings full of machines that cost billions of dollars.

Their existence means it's possible to reach such efficiency at manufacturing and heat dissipation. It means further progress, like the one we had until now, is possible. It seems that we currently are several orders of magnitude away from what human brains achieve.

Of course human flight doesn't use flapping wings nor do computers have to work 100% like brains. But they should at least be as efficient until we give up and say "no improvements are possible".

It's possible with organic chemistry, but what about inorganic chemistry? It's possible with carbon based chemistry, but what about silicon-based chemistry? I'm not convinced that it's impossible to build a human-mind-level consciousness in silicon, but I'm skeptical.

Human brains actually move real stuff around for their computation [1]. Like are like abacusses or some complex clock. Silicon based computers move electrons around which are much smaller and have much less weight. However, you might be right that silicon based computers won't allow further improvements.

So maybe one day we will build computers based on organic chemistry. If they are cheaper and more efficient and work as well as silicon based computers, there is certainly a case to build them.

[1]: https://en.wikipedia.org/wiki/Action_potential

Human brains are the most efficient at some problems. This is trivial to demonstrate: try multiplying two 10 digit numbers in your head. Now pick up a calculator and do the same in a fraction of time, using a fraction of power, and with perfect accuracy.

Computers are better than humans at some problems but I'd argue that this is because they can use their available computational powress much better than human brains can, even if it's much less. Computers are still much worse at driving cars for example and their power consumption is a real problem. If you could have done more computations for the same power budget, or with a power budget of tens of kilowatts, the problem might be solvable.

What about sub-atomic particles? It sounds artificially limiting, academically speaking, to say that a single atom is as small as we can go.

yeah, after nanometers we've still got picometers, femtometers, attometers, angstrom, fermi ... lots of room before we hit planck! intel, get your shit together!

The fact that the brain exists says absolutely nothing about the future of chip manufacturing. It is on a completely different plane than Moore's law: neurons are only as small as .004mm at minimum. What you're suggesting isn't an improvement, it's starting over completely from scratch.

But just to humor you, today's best processors have a transistor density of roughly 25M transistors per mm^2. The human brain has a neuron density of 14K neurons per mm^3. And of course, the human brain has a volume 250x the average desktop processor.

Depends on the brain doesn't it? Kim Peek and Srinivasa Ramanujan would indicate there is no fundamental limitation to the architecture of the human brain that prevents "fast math".

Should we measure CMOS and silicon's overall usefulness as a medium of CPU design based on the performance of a Z80?

the latency for operations in/ out of the brain is just incomparable between biological systems and cpus.

Not to mention the same issue with time needed for new algorithms/software to be generated, run, and limitations/difficulties on interfaces.

They're different systems, optimised for different things.

CPUs look superior if you ignore the problems using/requiring the huge benefits of biological and analogue systems the brain has, and the brain generally looks superior if you forget time, logic and deductive system implementations, inspection, standardisation, commoditization, etc.

obviously that's extremely simplified, but that's what one can fit in a HN post.

that also doesn't mean we can't learn and gain benefits for both by taking inspiration from the other, but like most things in life its far more complex than one outperforming the other or the two even being directly analogous.

I think you may have underestimated quite how fast computers are at maths. A $5 Raspberry Pi Zero, at 24 gigaflops, beats the combined performance of every American (continent, not country) combined, even if those humans were all calculating at the speed of the current world record holder, Takahiro Asano, who correctly added 15 sets of 3-digit numbers in 1.68 seconds.

(Against merely “average” human performance, a π0 would beat the entire planet by a factor of six).

That’s because humans have to decode written or verbal instructions, process them in a network that actually understands the concepts behind counting, numbers etc, and then translate the results into written or verbal language.

Wiring up some neurons to perform binary arithmetic could be much more efficient. But there was never an evolutionary reason to do so.

Well, yes — I’d call that a ‘fundamental limitation to the architecture of the human brain that prevents "fast math"’.

A fundamental limitation of the human brain, but not necessarily a fundamental limitation of organic neural networks.

I wonder whether a human calculating a math problems is not unlike a computer simulating another computer. Observing the hand-eye coordination of top athletes, it's pretty clear that our brains do some amazing computations, and fast. In fractions of a second, we can turn a pattern of photons into concepts, discover relationships between those concepts, track changes over time, and project all of that out some distance into the future. In other words, aren't most of our computational capabilities buried in the unconscious mind?

> not unlike a computer simulating another computer

It's more like this: http://dwarffortresswiki.org/index.php/User:BaronW#The_Almig... . Not just horrendously inefficient, but also highly at odds with what the brain is designed to do.

That might be a better way of making my own point than what I actually wrote. You’re right, explicit conscious maths isn’t what our brain architecture evolved for.

Of course, in terms of raw speed, my laptop can learn to read handwritten digits from only the examples in the SciKit-learn python module in 0.225 seconds [1], a bit less than the time it takes a human visual system to go from “some photons have hit the retina” to “there is a thought now, and that thought is ‘three’.” — the architecture of the AI is nowhere near as example-efficient as the architecture of a human brain, and it is only winning by the absurd performance difference of the hardware [2].

[1] https://kitsunesoftware.wordpress.com/2018/03/16/speed-of-ma...

[2] https://kitsunesoftware.wordpress.com/2017/11/26/you-wont-be...

You're making a category error. Human brains aren't computers. They do not "compute" at all.

Heat is already a limitation for going more 3D. At best you can slightly reduce latency, but with real tradeoffs.

To be fair growing a brain is also a VERY expensive process. To do it properly takes decades and can cost a lot of money.

Luckily it looks like humanity has finally figured out how to do it at scale /sarcasm/

The latest version of humanity isn't producing as many brains as it should.

> To be fair growing a brain is also a VERY expensive process.

Nah, you are born with 3 times the number of synapses you will have when you are an adult. The expensive part is training that network, not growing it.

Except heat doesn't cooperate

Brains have the benefit of an elaborate microscopic network of coolant pipes that double for power delivery

Exactly, and I've kept an eye on Nature Materials and other journals that cover advances in semiconductor cooling but so far there hasn't been anything very exciting. I was hoping the ultra thin silicon guys would use diamond as a substrate for its thermal properties (sort of like silicon on sapphire but silicon on diamond instead) but apparently not a thing yet.

The brain also runs at a much lower frequency than a CPU. A neuron can only fire about 200 times per second, or 200Hz. This is important because power consumption is directly proportional to frequency squared.

power and frequency are linearly related, it is the voltage that scales power in a quadratic order.

The power dissipation problem of current chips (performance is almost always power/thermal limited) is correlated with 2D structures. Communication is one giant problem: 3D structures could allow for far better colocation of connected compute elements, so we might not have to spend most of our power budget feeding data in and out of compute clusters. Scaling in 2D does make this problem easier (in some ways!), much as adding a 3rd dimension would too.

The power dissipation problem of current chips (performance is almost always power/thermal limited) is correlated with 2D structures

Graphene thermal pads have really phenomenal lateral thermal conductivity. 3D computer chips could be constructed of alternating silicon and thermal graphene layers, perhaps conducting to a central core, which could be made of copper or could even contain water or a heat pipe.

https://www.youtube.com/watch?v=YpphKzmDiJM

with the current semiconductor technology I doubt that 3D is an option for the foreseeable future. You would need a whole new implantation (or doping) technology for that

Stacks of thin silicon chips of 100 layers or so with microfluidic channels between the layers, or in the layers, sounds like the way of the future to me.

> Right now, CPUs are mostly two dimensional with only a few layers stacked on top compared to the extent in the other two dimensions.

Moving away from that would still be increasing transistor density and thus would not be rewriting or beating moore's law, it would still be moore's law applying.

I, for one, welcome the end of our Moore's Law overlords.

This race to increase various aspects of performance has long let the fields of architecture and the software on top of it be lazy in their optimizations. When there's no more juice in that fruit to squeeze, I know there is much more to be found elsewhere.

/I can't wait for my non-von neumann SoC FPGA hybrid to use for...I'll figure it out then I'm sure.

It would at least be nice to freeze the hardware for 10 years and give software time to learn how to fully extract what we have. Most large scale software ends up a factor of 10x to 1000x slower than it needs to be.

Most software isn't about performance. And when it does become a problem, then the goal is just to fix that quickly to the point of being good enough, even if it's a hack, like caching a very inefficient operation. I don't see anything wrong with that. Software is written to solve a problem, not to run as efficiently as possible.

Often people, especially computer engineers, focus on the machines. They think, "By doing this, the machine will run faster. By doing this, the machine will run more effectively. By doing this, the machine will something something something." They are focusing on machines. But in fact we need to focus on humans, on how humans care about doing programming or operating the application of the machines. We are the masters. They are the slaves. --Yukihiro Matsumoto

I'm very mean to my slaves. I have absolutely no issue leaving a rack of machines to slave away all weekend while I enjoy myself as opposed to spending that same weekend myself making it more efficient. (e.g. training models)

I don't care if a CPU is running at 80% consuming more electricity when I could work for a week and make it do the same job at idle. It's an inefficient use of my time. That same week's worth of effort could be spent delivering more value than the $10k server + $500/year operating cost. (I'm paid less than that, but I deliver more value to my company than my paycheck.) Also, risk. If it works now, it may not after I spend a week mucking with it.

> I don't care if a CPU is running at 80% consuming more electricity when I could work for a week and make it do the same job at idle. It's an inefficient use of my time. That same week's worth of effort could be spent delivering more value than the $10k server + $500/year operating cost. (I'm paid less than that, but I deliver more value to my company than my paycheck.) Also, risk. If it works now, it may not after I spend a week mucking with it.

That attitude is perfectly fine and reasonable until people start applying it to user-facing software. That's how you get software that wastes millions of user hours and immense amounts of energy just so some programmer could "deliver more value" and, sadly, that applies to an alarming proportion of modern software.

Those same users are probably benefitting from the other thing I did that delivered value to my company. It has value for the company because it has value, directly or indirectly, for the customer. Eventually, it'll sort itself out with competition.

Ultimately, my work is dictated by what the customers whine about. If they make more noise over missing feature X than slow feature Y, X gets my time.

> Those same users are probably benefitting from the other thing I did that delivered value to my company. It has value for the company because it has value, directly or indirectly, for the customer. Eventually, it'll sort itself out with competition.

This kind of oversimplified faith in the idealized notion of the free market completely ignores the myriad of externalities that systematically distort software competition. General arguments aside, it suffices to go through almost any set of software niches to note that the dominant products in most of them are dominant for reasons that have nothing whatsoever (and are frequently inversely correlated with) any reasonable sense of technical merit or even economic value produced (for the users) in the present (though many of them were once upon a time arguably the best in their respective niche).

> Ultimately, my work is dictated by what the customers whine about. If they make more noise over missing feature X than slow feature Y, X gets my time.

Non-technical users rarely articulate performance-related concerns, but that doesn't mean they don't have them. In my experience, if you have prolonged free form conversations with them, they will often voice impressions of the software in question getting slower and less reliable overall, whereas targeted questionnaires and focus groups will tend to laser in on specific features simply because those are much easier to articulate. Also, bad software is so common by now that some people can't even imagine that there isn't actually any sane technical reason for, say, Photoshop to take almost 10 seconds to open a simple image file. That doesn't mean you're not wasting their time, money and energy.

You see, that's the difference between a craftsman and a corporate drone.

There are some minor and major things wrong with this attitude:

1. it wastes energy, increasingly a non trivial share of total energy use by humans

2. it often manifests itself as the user (me) having to wait for things that should be instant.

If the latter never happened, I would be mostly on board with what you are saying. I strongly wish we could improve the state of software such that you could achieve the same productivity we have now, with only a factor of 2 to 10 slowdown vs optimum instead of the 10x to 1000x we accept now. I think that should be possible. As a start, we can make first class jits for languages that are usually interpreted( python / ruby ) and possible, come up with statically typed languages that people find as nice to use as they find dynamically typed ones. I think this should be possible with techniques like type inference. Doing those two things can improve the efficiency of a lot of code without any downsides imho.

https://news.ycombinator.com/item?id=9224

The problem is that after those 10 years new technology could differ so much that all efforts put in optimization would be pointless.

Things like that already happened - for example slow HDDs are now replaced by very fast SSDs. There is Intel Optane that is only 1 order of magnitude slower than RAM and Intel Opate DIMM is closing gap even further.

Me too. As long as reducing feature size is possible, it'll never make sense to try anything really new. Nobody would have ever built a car if horses halved in price and doubled in performance every two years.

No, but they would eventually have $10 faster-than-light horses, which would be a lot better than a car.

I think you're romanticizing a bad thing.

I think it’s fine to look on the bright side, as the “bad thing“ in this case seems to be something we’re stuck with.

It simply isn’t possible to make transistors infinitesimally small (just as a horse’s metabolism could never keep up with the relativistic muscle movements required for faster than light travel.)

Sure, that sucks. But realistically, the only way to progress is to design better architectures, or think of new computing paradigms, or deepen our understanding of the laws of nature.

All of those, if realized, could be very good things.

I sometimes get bummed out about the end of Moore’s law, but I try to remind myself to be grateful and appreciate what we have. Computers are wonderful, and it’s amazing that we have them at all: the universe didn’t owe us computing; we figured it out in order to solve problems. Now we have another big problem to solve.

Ultimately, we will see huge investments in pushing the boundaries of computing. Whether it’s biological, quantum, or just massively parallel, I’m excited to see what we can come up with in the years ahead.

Do you think that "FLOPS/OPS/whatever per year" will slow down or speed up now that Moore's law is ending? Ie do you think that the hardware/architecture/software combo will accelerate more now, or less?

If you think it will accelerate more while hardware gains drop to zero, why were we leaving these absolutely massive gains on the table? This means that, if a new processor gave us double the speed, we could get more than double just from software alone and get the hardware doubling on top of that.

If you think it will decelerate, why are you happy Moore's Law ends? We're worse off than we were before, even though we aren't "lazy" now.

This approach seems to me a bit like saying "Oh I'm happy I lost my job, now I can finally make money by looking for pennies on the ground all day, which the job made me too lazy to do".

By ending, they are absolutely referring to it slowing down.

It probably means we're worse off, in that spending on software engineering to get gains rather than expecting it out of new hardware processes will almost certainly be more expensive. On the other hand, software gains give us more computing with the same energy which is a good thing for the planet.

Some of us liked looking for pennies. Watching the world burn so I get lauded rather than ridiculed for micro optimizations is a small price to pay.

It is also possible we are in a local optimum and there are huge performance or efficiency or other kind of gains available, but to realize them investments in radically different architectures need to be made, and those are not economical while existing approaches can be improved.

This is exactly my perspective. Existing ISAs have a de facto stranglehold on the mindset around computer architecture because they could provide a consistent foundation while the underlying implementation got better and better faster than anyone could build something on top of it.

There have been plenty of small experiments in other architectures but never enough of a mindshare investment to really evaluate viability and performance of alternatives. There are costs to a more diversified architecture eco-system, sure, but I fully believe we will find more value in that space than we have so far taken advantage of.

ASIC implementation of a PLC, anyone?

OPS might speed up, while FLOPS will slow down :)

By FLOPS I mean the classical FP64, while OPS can be any bit width, even 1 bit.

> I, for one, welcome the end of our Moore's Law overlords.

I think if you fully understood the implications of this, you wouldn't be so sanguine.

We're already seeing the effects of a stalling out of process improvements.

Starting a couple years ago, there were articles by enthusiast PC builders who were buying used Xeon processors on eBay, and building new gaming PCs around them, with comparable performance to the i7's available at the time. Just to put some round numbers on that, you had people spending $100 or so on a used CPU (none of that money goes directly to Intel) vs. spending nearly $1000 on a brand new i7.

You can find other instances where a company is competing with their own older, used products... it isn't fun for the company in question.

What's going to happen to Intel's stock price when that sort of thing becomes more common? The price of their stock (and all other tech companies) is based on future value, rather than other companies that pay dividends but otherwise aren't expected to expand significantly.

If investors come to accept the fact that Intel can't really make a chip next year (that will drive hardware and software sales for the entire PC industry), better than last year's chip, their stock will crash. As will all the others in the tech industry.

New hardware (with more capabilities at a reasonable price) drives new software / new applications. If next year's phone isn't better than last year's phone (assuming you can replace the battery that is losing capacity), that will kill valuations in the mobile segment.

Some tech like stacking higher in 3-D on the chip will help applications like NAND Flash storage that aren't heat-dissipation limited. But even there the cost per bit stored will won't be going down at nearly the rate like we're used to seeing.

A tech crash across the board will tank the entire world economy as well, because much of the productivity improvements are based on computer technology.

This is a massive, massive upheaval we're looking at. Unless we switch to new technology that can keep the productivity party going for a while...

"What Andy Giveth Bill Taketh". Software still can solve problems and deliver value.

Oh, we've a long backlog of software improvements.

The question is, what drives stock values? What drives the entire industry?

What does it matter to the economy if Intel's shares are low or high?

It won't be just Intel, but the entire semiconductor industry, and everything that depends on it.

I was mentioning Intel because they have been a leader in semiconductor process technology, often the first to ship product at a new process node.

Product differentiation using specialist architectures will be no problem for 10 -15 years. After that; software.

"After all, there aren’t many numbers left between 5 and 0."

Math purists may roll their eyes at this statement and clearly even if you are obsessed with integers you can just switch to picometers, but I wonder if it is time to start counting atoms. The diameter of a silicon atom is about 0.21 nanometers, so a 5nm process is dealing with features only about 20-30 atoms wide.

(edit: found more accurate diameter number)

Wow, this puts things in perspective for me. I knew 5nm was small but I didn't really realize how small until we started getting into "I-have-more-dollars-in-my -wallet-than-this-thing-is-atoms-wide" territory.

One thing I’m always amazed with - and it’s partly due to my lack of understanding of physics at this scale - is that when a processor with components on this scale is dropped a few feet, they aren’t reduced to smithereens. It’s hard for me to contemplate that whatever is holding these few atoms together is so resilient.

If you find that amazing, contemplate for a moment how nuclear fission and fusion work and the amounts of energy involved there. Apart from a few comparatively exotic things (black holes, neutron stars, etc.), some of the largest force densities that exist are all around us in atomic nuclei (of course, those forces also decay very quickly with distance, but that's a different matter).

The "components on this scale" are pretty light so there's not much force when you accelerate them. If a drop doesn't crack the silicon wafer it's not going to shift the bits and bobs on it.

Not sure how correct this is, but I think of silicon as basically a small metallic rock.

They are made up of a solid lattice, the components are formed by differential doping. There is some migration, but it is far too slow to have an effect over the lifespan of our electronics.

Physical inertia is meaningless over other forces involved in the lattice.

The node label (e.g. "5 nm") no longer represents a physical dimension and hasn't for some time -- it's simply a continuation of a naming convention. In fact, while we used to scale down transistors by simply reducing the gate width (where the naming convention originated) we now take far more elaborate steps to do so.

Fin width is around the same size as the node name, but yeah gate length at 7nm is nowhere near 7nm.

What ways to get faster cpu remain after this? Nanoscale vacuum-channel transistors [1] seem promising as they can work at terahertz frequencies but they do not look anywhere close to production.

[1] https://en.m.wikipedia.org/wiki/Nanoscale_vacuum-channel_tra...

I remember hearing about research into germanium as a replacement to silicon or compound III-V semiconductors [1]

Nanoscale devices seem like an interesting development, but all of these technologies seem like they're a decade or more away from mainstream use.

[1] http://www.sandia.gov/%7Ejytsao/WCS.pdf

I think you may mean gallium based compounds, not germanium. Germanium semiconductors are only good for sounding like Hendrix. The RF biz debated sticking with Si or moving to "wide bandgap" materials like Gallium Arsenide (GaAs) or Gallium Nitride (GaN), because the wider bandgap means the devices have much higher breakdown voltages and better power density. That means you can run them at a higher baseband and lower power without a die shrink.

It's my understanding they're important for 5G devices, and they've been in the wild for awhile. But the kinds of circuit they implement are like demodulators and filters, not so much something with the transistor density of a microprocessor. I think the fab processes are well more than a decade away from reaching the die sizes of contemporary Si, even if hypothetically they could run at ridiculous clock rates and low TDP.

"Germanium semiconductors are only good for sounding like Hendrix"

You say that like its a bad thing.

SiGe is a serious competitor to GaAs and InP. It doesn't typically have the power density of Ga derived microcircuits, but has good channel noise performance plus linearity and can operate at higher frequencies than GaAs as long as you don't need high power density. GaN has amazing power characteristics but has all kinds of issues with signal purity that are still being worked out. Eventually, I believe GaN will eat everything's RF lunch just out of sheer power efficiency gains with the last hold out being cell tower base stations, where spectral purity trumps efficiency by a mile. InP gets used into the 100GHz+ territory.

These processes are mostly larger like you say, but they are also used in high end data converters (ADCs and DACs) and some processes are in the mid to low 10's of nanometers now. Which is almost exactly where silicon was a decade ago.

Silicon of course isn't used because of it's performance, but rather, because silicon is the easiest of the semi-conductors to manufacture and get good yields out of. It is also comparatively very low cost. These factors have led to the lion's share of research dollars being spent there as well.

Direct S/D tunneling becomes a problem with high carrier mobility semiconductors like III-V group materials and germanium.

I think moore's law has spoiled a generation on what technology can do. It is extremely ridiculous that lithography has gotten so much better in the last few decades, and IMO it has resulted in a lot of people thinking that exponentials are a) common, and b) automatic. Neither of these are true.

There are few, if any, other exponentials in the history of humanity that worked as consistently for as long with as massive of impacts on technology, and TBH I'm doubtful that we will see anything like it again.

Moore's law worked because loads of engineers and scientists made it their life's work, but everyone watching seems to think it just happened.

There was one magical exponential that changed the world, and it's coming to an end.

There's certainly a lot that can be done in architecture and software, but most applications probably aren't worth the effort.

I recommend watching the Turing Award speech by Patterson and Hennessey which covers this topic.

Spoiler: the way out is "Domain Specific Architectures".

https://www.youtube.com/watch?v=3LVeEjsn8Ts

Which is sad, in a way. There was something amazing about a general purpose computer that next generations may not be able to appreciate.

General purpose chips aren't going to disappear, not even slightly. They're just going to offload specific work to circuitry that it's best suited for.

I regret we're about to hit the wall but I have long though that we're going to start moving back to analogue, and probabilistic algorithms, and in one way, having half-width floats is a step in that direction. I expect it to go further.

(All this is my opinion and may be wrong).

Sure, but you might end up with computers for video editing, computers for DNA sequencing, etc.

And we are already there for some workloads. AI and crypto mining come to mind.

Regardless, buying a fast computer that could do anything was really fun.

I think you will see SOC's that have multiple specialised architectures, which will be lit up intelligently by the CPU depending on what the workload is.

> They're just going to offload specific work to circuitry that it's best suited for.

Like the Amiga? That actually worked out pretty well. I look forward to it.

Pow’s law at work here for me

https://en.m.wikipedia.org/wiki/Poe%27s_law

It was all a marketing illusion to sell pentium chips and windows licenses. Bittersweet to see it go, as it was a great time to start being a technician and I certainly personally benefited from it. We're on the way back to using the Right Tool For The Job now, and that's probably better for everyone. Workloads are unique, just like people.

That's a bit like saying it's sad that cpus added media instructions, or that the gpu was invented. Dedicating silicon to fixed tasks or task specific designs can lead to massive performance increases and in no way means the death of general purpose compute.

And it is, in a way. At least to me.

Would you rather have a burnt on silicon algorithm for video/encryption decoding or a 10X faster general purpose CPU that can do that and everything else?

I think you underestimate just how much faster and more energy efficient hardware acceleration can make certain tasks, the savings can be far in excess of 10x.

You may underestimate the exponential function :)

We where doubling our performance every year or two. For decades.

That doubling applies to any sillicon circuit, not just CPUs.

Of course it does, and your point is?

We were also watching 160x240 video clips.

The problem with this is the likely fragmentation of the programming languages and tool support.

Eg GPUs have been around for ~25 years and there's still no decent way to portably program them. We have Metal, Direct3D, OpenGL, OpenCL, Cuda, Vulkan, ... all of which have small market shares. It's markedly worse than microprocessors at 25.

(Not even getting into quality issues like driver quality problems and the closed&proprietary nature of most of the SW stacks).

The end of Moore's law doesn't bother me too much from a practical point of view. My phone and computer are about as fast as I need them to be. Why it does bother me, though, is how it's an inescapable reminder that there are hard limits to everything, and sooner or later we're going to hit them. The fact we're hitting this limit so fast is damaging my hope for a whiz-bang sci-fi future. I didn't expect to be alive when we invented strong AI and perfect 3D projection systems and retinal display screens and all the rest, but I did hope we eventually would. Slamming into the end of per-volume and per-watt performance before the year 2100 makes me think maybe those things will never happen at all.

It's worth noting that we hit the wall only with respect to silicon-based transistor-packing wafer technology. There are other alternatives (ex. SiGe instead of just Si, memristors, photon instead of electron circuits, etc.) that could allow significant increases in compute capacity. They just haven't been explored or invested in to a large extent because it's been relatively cheap with proven ROI to follow traditional manufacturing paths to their maximum. Other technologies like Li-ion batteries are in the same boat - there are big advancements yet to be made in the broader space in general, but the specific processes, chemistries, etc. being used today are nearing their physical limits.

I'm personally still hopeful for a sci-fi future, just based on a series of ramps and plateaus as we progress through the limits of each specific technology.

Don't despair. We still have economic, energy, and material science boundaries to smash.

Just imagine your current phone costing pennies, compostable, and easily chargeable anywhere. Imagine never again worrying about file size or storage capacity.

We've got a long way to go and I'm very excited about what the future holds.

Yeah, but the ways I want my devices to improve are more "cultural" than technological. I want easily-rooted phones with expandable storage, 3.5mm jacks, and screens that aren't five times taller than they are wide. I want laptops with non-chiclet keyboards and real mouse buttons instead of clickpads where I can't tell where the right click starts.

Well if 'improvements' stop, then you'll start getting those things.

Take Librem, their biggest problem is they don't have millions of $ to throw at the problem, so everything takes longer, so you end up with a flagship phone, of 5 years ago.

Now if the state of the art stays the same for 5 years, a 5 year dev cycle doesn't matter, so these smaller companies can compete.

Or take current notebooks for example. Thinner and lighter devices in the same form factors were the trend for years, but now that that's hit the point where thin compromises every other feature, there are designs being shown, especially in high-end gaming notebooks, that do more radical things: shift the keyboard downwards for thermal efficiency or to add a second screen; or a screen that detaches and elevates for better ergonomics.

There's a lot of room for electronics to become better physical objects, and for the software to mature with those objects. Moore's law also meant that the software got replaced with new software every four years or so.

It is a strange and sad feeling to come to the realization that everything is finite.

With possible exceptions for space, time, prime numbers and the like.

For all we know space is finite and there is a countable number of stars in the sky.

Yeah hard to know really on that one.

The end of Moore's law probably won't stop strong AI. We are already getting to processing power comparable with the brain.

A normal human brain has thirteen thousand times as many synapses as AMD's Epyc has transistors, each capable of far more complex behavior than a single transistor. It's going to be a while.

What clock speed do the human synapses run at though?

Humans don't have a central clock therefore this is a meaningless question. Even if your neurons only fire 100 times per second they can be interleaved. Therefore two neurons firing at that speed could be offset by half a clock cycle and therefore the global clock rate is 200Hz. As you add more neurons the global clock rate rises. Therefore the clockrate is anywhere between 100Hz to 10THz depending on how you count it.

I was thinking of Hans Moravec's calculations based on being able to perform the same function such as visual recognition at the same rate and resolution as the brain. He figured about 100 terraflops. https://jetpress.org/volume1/moravec.htm

Which you can now kind of get in a $7k Nvidia GPU (the Tesla v100). Not saying that has all you need for AGI but the hardware is getting there.

The brain doesn’t even do visual recognition though. It does vision-directed action.

I was a bit unclear - I was thinking of the bit in his essay discussing the image processing done by the retina.

$100million per lithography machine!

A bit of a side topic, but can you imagine if there were fewer people in the world? Computing power would never improve because it would be too expensive.

I wonder what other technology could be done, if only there were more people (i.e. larger market), but is too expensive right now.

People always like to talk about the downsides of population, but the high tech world we live in now could not exist otherwise.

And there wouldn't be enough smart people.

These lithography machines are completely wild: the part that fails the most spits out a tiny droplet of molten tin. That's hit with a laser, and the excited electrons emit what's called extreme ultraviolet (EUV) which is very close to the softest of X-rays. Most of the power released is wasted bouncing off a series of mirrors, and then wild stuff happens as these energetic photons hit the semi-optional mask cover (pellicle), mask, and then chip: https://en.wikipedia.org/wiki/EUV_lithography

"After all, there aren’t many numbers left between 5 and 0." For some reasons, I found this sentence very funny.

Probably because it's both patently untrue and still informative in a way. It would be more precisely stated as "there aren't many atoms left between 5(nm) and 0". If my math is right, 5nm is only a width of about 25 silicon atoms.

Yes you would think ieee would know there are uncountably many numbers between 0 and 5.

IEEE surely does know. I'm sure the point is to imply the marketing ridiculousness. After 1nm they'll need a new vanity metric.

1nm is about 5 atoms so I guess the next metric would be the number of atoms. Single atom features would be kind of cool but not buildable with present technology.

No, there aren't uncountably many relevant numbers between 0 and 5, because you have to make your chip out of atoms.

It is also patently untrue because "5nm" label has nothing to do with reality.

It isn't so far off the fin width. And practically, it isn't a terrible scaled density metric. Increasingly bad, but not terribly so.

On the 5nm: it is more useful to think about scaling minima in terms of unit cells. For silicon, 5nm is about 10unit cells, where silicon behaves approximately.bulk-like (like silicon).

When you make it thinner, it is statistically no longer uniform.

The same logic applies everywhere: you need roghly 5nm of fin 5 nm of gate, and 5nm of space: fin pitch is limited at 20nm. For metals it is the same, but worse for resistance reasons, so pitch is limited to about 20-25nm. Gate is even worse.

Even though I think all of this is sort of clearly true, there is so much room for innovation elsewhere. I. Package, stacking, memory....there are 4-5 orders of magnitude out there in power-performance. In silicon!

Plotting the cost per compute power over time gives a different image:

https://sv.wikipedia.org/wiki/Teknologisk_singularitet#/medi...

In this display, the show is far from over. It even start to look like over-exponential progress. The latest data points show GPUs instead of CPUs but this is most likly OK since it is the suitable technology for easy to parallelize workloads like machine learning applications.

Anyone know if there's a longer version of this chart which goes back farther in time?

https://spectrum.ieee.org/image/MzMwNzU1Ng.jpeg

Anyone care to confirm that Intel Custom Foundry is still up & running?

Do you mean at 14nm? I guess once they successfully migrate their actual production line to 10/7nm, whenever that might be..

But think of all the interesting work ahead! :)

I kind of welcome the death of the "just wait for hardware to improve" approach to optimization. I find computers interesting because the field is so fresh and you have to figure out so many things on your own and through communication with people that are still alive. Due to the death of Moore's Law, it's going to stay that way for some longer and the end result is going to be more varied.

Nice article with a bit of low-level details. We are indeed encountering the problem that light wavelength is now small enough for us! So using 15nm EUV for multiple patterns to get 5nm. Only two foundries left in the entire world which can do such process (Intel and others are years behind).

> Intel and others are years behind

Not clear anyone other than Intel is trying, but Intel's situation is much more complicated. They tried for a purely optical for its life "10nm" node more aggressive than TSMC's initial "7nm" node and catastrophically failed. There's clearly some top level management issues there, a big problem for Intel for decades, and it's extremely worrying they let them destroy a generation of their crown jewel.

But that doesn't tell us much about their "7nm" node very roughly equivalent to Samsung and TSMC's "5nm" nodes, except those companies have a lot more real world EUV experience, but it's not always good to be a pioneer. Intel could conceivably get back in the game in about the same time frame as these two companies exit "5nm" risk production, we just don't know. All we know is that they're buying EUV machines and installing them in multiple fabs.

Stupid question. Could trinary processors play a role here?

I keep hearing people talking about them and their benefits over que binary system.

What operations do you perceive would be faster or more suitable for ternary processors

>I keep hearing people talking about them and their benefits over que binary system.

Who are these people?

Reply on news.ycombinator.com