> Another possibility that has long been on my personal list of “future articles to write” is that the future of computing may look more like used cars. If there is little meaningful difference between a chip manufactured in 2035 and a chip from 2065, then buying a still-functional 30-year-old computer may be a much better deal than it is today. If there is less of a need to buy a new computer every few years, then investing a larger amount upfront may make sense – buying a $10,000 computer rather than a $1,000 computer, and just keeping it for much longer or reselling it later for an upgraded model.
This seems improbable.
50-year-old technology works because 50 years ago, transistors were micron-scale.
Nanometer-scale nodes wear out much more quickly. Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
One of my concerns is we're reaching a point where the loss of a fab due to a crisis -- war, natural disaster, etc. -- may cause systemic collapse. You can plot lifespan of chips versus time to bring a new fab online. Those lines are just around the crossing point; modern electronics would start to fail before we could produce more.
> Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
That statement absolutely needs a source. Is "usage" 100% load 24/7? What is the failure rate after 7 years? Are the failures unrepairable, i.e. not just a broken fan?
Yeah, chip aging is at max temperature, max current, and worst process corner. And it's nonlinear so running at <10% duty cycle could reduce aging to almost nothing.
Every now and then, I get a heartfelt chuckle from HN.
By 'Modern' they must mean latest generation, so we'll have to wait and see. I was imagining not using an RTX 5090 for 7 years and find it doesn't work, or one used 24x7 for 3 years then failing.
I’ve never heard of this and I was an Ethereum miner. We pushed the cards as hard as they would go and they seemed fine after. As long as the fan was still going they were good.
So Intel used to claim a 100,000+ hour life time on their chips. They didnt actually test them to this because that is 11.4 years. But it was basically saying, these things will last at full speed way beyond any reasonable life time. Many chip could probably go way beyond that.
I think it was about 15 years back they stopped saying that. Once we passed the 28nm mark it started to become apparent that they couldnt really state that.
It makes sense, as parts get smaller they will get more fragile from general usage.
With your GPUs yeah they are probably still fine but they could already be half way through their life time, you wouldnt know it until failure point. Add in the silicon lotto and it gets more complicated.
One thing to realize is the lifetime is a statistical thing.
I design chips in modern tech nodes (currently using 2nm). What we get feom the fab is a statistical model of device failure modes. Aging is one of them. When transistors gradually age they get slower sue to increased threshold voltage. This eventually causes failure at a point where timing is tight. When will it happen varies greatly sue to initial conditions, exact conditions the chip was in(temp, vdd, number of on-off cycles, even the workload). After an agong failure the chip will still work if the clock freq is reduced. There are aging monitors on-chip sometimes which try to catch it early and scale down the clock.
There are catastrophic failures too, like gate insulator breakdown, electromigration or mechanical failures of IO interconnect. The last one is orders of magnitude more likely than anything else these days.
For mining, If a GPU was failing in such a way that it was giving completely wrong output for functions during mining, that'd only be visible as a lower success hash-rate which you might not even notice unless you did periodic testing of known-target hashes.
For graphics, the same defect could be severe enough to completely render the GPU useless.
> Nanometer-scale nodes wear out much more quickly. Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
I recently bought a new MacBook, my previous one having lasted me for over 10 years. The big thing that pushed me to finally upgrade wasn’t hardware (which as far as I could tell had no major issues), it was the fact that it couldn’t run latest macOS, and software support for the old version it could run was increasingly going away.
The battery and keyboard had been replaced, but (AFAIK) the logic board was still the original
> it couldn’t run latest macOS, and software support for the old version it could run was increasingly going away.
which is very annoying, as none of the newer OS versions has anything that warrants dumping hardware to buy brand new to run them with! With the exception of security upgrades, which i find dubious for a company to stop creating (as they would need to do so for their newer OS versions just as well, so the cost of maintaining security patches ought to not be much, if at all), it is definitely more likely to be a dark-pattern to force hardware upgrades.
That's not just a dark pattern, it's the logical conclusion to Apple's entire business model. It's what you get for relying on the proprietary OS supplied by a hardware manufacturer. It's why Asahi Linux is so important.
"regularly" is doing a lot of work here. When Linux drops hardware support, we are talking about ancient hardware. An example of a regular drop: Linux 6.15 just a month ago dropped support for 486 (from 1989)!
Open source software drops hardware support only when there are nobody left who volunteers to support that hardware. When does this happen? It happens when there are not enough users left of that hardware.
As long as there are enough users of some hardware, free software will support it, because the users of that hardware want it to.
Depending on how much has changed in the interval, backporting security fixes can be completely trivial, very difficult, or anywhere in between. There may not even be a fix to backport, as not all vulnerabilities are still present in the latest release.
You mean besides the fact that they completely transitioned to a new processors and some of the new features use hardware that is only available on their ARM chips?
Also he said that software from third parties also don’t support the older OS so even if Apple did provide security updates, he would still be in the same place.
I've got 3 Macbooks from 2008, 2012, and 2013. Apple dropped MacOS support years ago. They all run the latest Fedora Linux version with no problems.
The screen on the MacBookPro10,2 is 2560x1600 which is still higher than a lot of brand new laptops. The latest version it will run is 10.15 from 2019. I know Apple switched to ARM but most people don't need a new faster computer. I stopped buying Apple computers because I want my computer supported more than 6 years.
I do have 3 newer computers but these old Macbooks are kept at various relative's houses for when I visit and wnat my own Linux machine. They have no problems running a web browser and watching videos so why replace them?
> Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
I seriously doubt this is true. The venerable GTX 1060 came out 9 years ago, and still sees fairly widespread use based on the Steam hardware survey. According to you, many (most?) of those cards should have given out years ago.
This is just untrue, and you’ve provided no citation, either.
The silicon gates in GPUs just don’t wear out like that, not at that timescale. The only thing that sort of does is SSDs (and that’s a write limit, which has existed for decades, not a new thing).
Electromigration tends to get worse with small sizes but also higher voltage and temperatures. I could see a GPU wearing out that quickly if it were overclocked enough, but stock consumer GPUs will last much longer than that.
since electromigration is basically a matter of long, high-current interconnect, I guess I have been assuming it's merely designed around. By, for instance, having hundreds of power and ground pins, implying quite a robust on-chip distribution mesh, rather than a few high-current runs.
Wouldnt it depend on work loads? My GPU that kicks into high gear for maybe 2-3 hours a week will probably do decades of use before chip degradation kicks in. The power capacitors will give out long before the silicon does.
But it someone is running an LLM 24 hours a day, might not go for as long.
We are flying blind, both on those claiming short life span and those who are not.
> One of my concerns is we're reaching a point where the loss of a fab due to a crisis -- war, natural disaster, etc. -- may cause systemic collapse.
This is absolutely ridiculous. Even if Taiwan sank today we really don't need those fabs for anything critical. i strongly suspect we could operate the entire supply chain actually necessary for human life with just z80s or some equivalent.
More parallelism. Less clock. More l1 cache per CPU and less disk stalls. Are there plenty of tricks in the sea, when clockspeed goes flat?
Dual ported TCAM memory isn't getting faster and we've got to 1,000,000 prefixes in the internet and ipv6 are 4 times bigger. Memory speed is a real issue.
There has been a relationship between die area and speed for much of time. That stopped being true. Perhaps it was only a side effect of shorter paths on chip.
What does economics in die area say about using die for highly repeatable simple memory forms, vs complex instruction/ALU? I would think there is a balance point between more cores, and faster cores with more cache and more lanes between them.
We're on 8 as a mixture of fast and slow, hot and cool, on almost all personal devices now. I routinely buy Intel instruction racks with 16 per CPU construct, and a TB of memory. At one remove, chips seem to be fulfilling my mission but I am not in GPU/TPU land. Although I note, FPGA/GPU/TPU seems like where a lot of the mental energy goes these days.
VLSI is a black art. Quantum effects are close enough to magic I feel confident saying few people today understand VLSI. I know I don't, I barely understand virtual memory. I ache for the days when a zero pointer was a valid address on a pdp11
I think its possible. When 128 bits came along, I wondered if the very long instruction word people were going to come back into the room. If we imagine mapping addressing into the world, how about making some rules about top bits addressing other nodes in the DC and see how "shared memory" works between distinct racks?
Moore’s law has been unsustainable for 20 years, I remember Pentium 4’s with 4ghz. But that hasn’t seemed to matter in terms of real day to day performance improvements. This article makes some great points about the scaling cost and reduced market opportunity for there to be more than 2 or 3 makers in the market, but that’s a trend we’ve seen in every market in the world, to be honest I’m surprised it took this long to get there.
As interesting as this breakdown of the current state of things is, it doesn’t tell us much we didn’t know or predict much about the future, and that’s the thing I most wanted to hear from an expert article on the subject, even if we can take it with a large pinch of salt.
>software is getting slower more rapidly than hardware is becoming faster.
>Wirth attributed the saying to Martin Reiser, who in the preface to his book on the Oberon System wrote: "The hope is that the progress in hardware will cure all software ills. However, a critical observer may observe that software manages to outgrow hardware in size and sluggishness."
I wish that there would be more instances of developments like to Mac OS X 10.6, where rather than new features, the software was simply optimized for a given CPU architecture, and the focus was on improving performance.
This doesn’t hold with my experience, I remember how slow my old machines used to be - especially the ones with spinning hard disks. Not to mention early smartphones… I’ve had computers in my youth I used to turn on then go and make a coffee. Now it’s practically instant.
Yeeeah dude you have some seriously rose tinted glasses on.
My windows machine can go from completely powered off to a big application like a web browser fully open and operational on maybe 10-15 seconds, and yes, fully powered off. No way I’m making coffee in that time. Sure, windows might run stuff in the background, but it’s so fast that stuff doesn’t get in the way of me using the thing at all.
As for phones, “forever” is a silly thing to say but even if I granted you that one off boot time, you have somewhat moved the goalposts as you talked about the act of using software in day to day usage. On my laptop I can genuinely open 10 apps, close 10 apps, and open 10 apps again all within 10 seconds, probably faster if I wasn’t constrained by literal click actions. My computers growing up would grind to a halt if you tried to run like 3 things at once!
There are always instances of inefficient software, but there was back in the day too, I remember generating a new map on civilisation 1 on my 286 and it taking 10 hours. There is no way you can possibly say that on average software is slower today than it used to be, that statement in general is just completely false.
> I remember generating a new map on civilisation 1 on my 286 and it taking 10 hours
Humm. If my glasses are rose tinted, yours are black tinted. I've played quite a lot of civ 1, even on an xt with 8088, and i'm pretty sure map generation did not take 10 hours.
Perhaps if you hacked yours to use larger maps...
I'm also pretty sure civ 1 got to the menu on my 8088 faster than Death Stranding 2 on my ps5. Although that might be my rose tinted glasses indeed. Civ 1 also didn't have a 29 Gb day one patch :)
The improvements you describe come down to one thing: solid state drives. If we had them in the 90s, your software would've been starting up just as fast (or even faster) then. Meanwhile, once the software did start, it ran faster and with fewer resources than software does today. It's not rose colored glasses: software today truly is worse on average than software 20 years ago.
The horrors of booting Windows 2000 on a Pentium 1.
But there is something to be said about prior optimizations in limited resources. A few months back a friend was restoring a G4 iMac and I was astounded at how snappy the whole thing felt once the OS had finally loaded.
We forget how snappy iTunes was while using only 30MB of RAM.
If you can combine the best of boths worlds, there could be something special there.
Modern OS and software are very bloated but it is astounding how fast the hardware crunches through all that. But it could be better.
Software optimized for a particular processor is today's tech. In the future we may see CPUs that alter themselves to better serve the needs of software. I can see a day where a "cpu" was actually a bunch of different processors connected via a FPGA so that the software could reconfigure the CPU on the fly.
The great thing about AI, it is finally a killer feature that is enjoyed and useful by every user worldwide. And the tech industry finally have an excuse to up sell 16GB as baseline, and perhaps even try to push 24 or 32GB Memory, along with GPU / NPU and CPU upgrade.
For users, a few hundred dollar extra ( on top of the original purchase ) is a such a small number compared to the productivity gain over the usage span of the computer.
AI alone not only increased the server hardware requirement but also user client requirement. It is basically the question everyone has been asking, what is after Smartphone? And to a degree it is AI. ( or LLM )
This will easily push the whole Semi-Conductor Industry forward all the way to 2032 ~ 2035. We will be at 8A or 6A by then.
PCIe 7? possibly PCIe 8? WiFi 9 which is a fixed version of WiFi 8. There are so many great Hardware improvement coming out all because of the demand of greater computing usage.
Software side has been rather boring TBH. I really like the phase Allan Kay uses to describe modern days software are "reinventing the flat tire".
I think the future of compute will look much like today!
Given the power and ubiquity of smart phones, most people don't need any other computer in their personal life. What can be done locally on a smart phone seems like it will be more constrained by battery life and physical size than anything else, and there will continue to be a mix of things that can run on-device and other more compute-hungry functions run in the cloud. I don't see smartphones being replaced or augmented by other devices like smart glasses - people want one device that does it all, and not one they wear on their face.
The same is somewhat true for business use too, especially if compute-heavy AI use becomes more widespread - some functions local and the heavy work done in AI-ready datacenters. I'm mildly surprised that there hasn't already been a greater shift away from local compute to things like Chromebooks (ubiquitous in schools) since it has so many benefits (cost, ease of management, reduced need to upgrade), but maybe it will still come if the need to rely on datacenter compute increases.
Even if we imagine futuristic power-sipping neuromorphic chips, I don't see that changing things very much other that increasing the scope of what can be done locally on a power budget.
> Given the power and ubiquity of smart phones, most people don't need any other computer in their personal life.
Maybe, but they need other form factors, too. Like a big screen and a keyboard.
And once you have those, you might as well through some compute in there: getting smartphone-like performance out of desktop-like components (or even laptop components) is pretty cheap, so might as well throw them in.
Chromebooks are fairly capable 'thin' clients. Their specs aren't that much worse than other laptops.
However all that being said, I mostly agree that at least with current style AI we will likely see a lot more data centre compute in the future. At least as long as hardware costs are a significant fraction of total costs: the 'cloud' can multiplex multiple users over expensive hardware, your local computer is mostly idle and waiting for your input. (And it has to be, if you want low latency.)
However, if costs shift so that electricity and not hardware is the main cost, we might see a return to local compute. Maybe.
> Maybe, but they need other form factors, too. Like a big screen and a keyboard.
It seems generational habits are changing. TV is dead, and younger generation seems ok watching things on small screens, even smartphone. Can always sync phone to parent's/hotel big screen TV if available.
God of war 2 was made on 300mhz cpu and 32mb of ram.
We haven't been bound by moore's law because we just waste computing power because programmers are expensive. No one is trying to optimize nowadays except in very niche places. And when push comes to shove we just start wasting slightly less. Like adding a JIT to a script language 15 years too late.
As someone who has met people whose entire job is to optimise games for consoles, this isn’t true. It’s ok to want to get the most out of hardware in terms of features too.
Can you point to a couple of games in the last few years that would make one exclaim "what kind of forbidden magic they did so they squeeze so much performance out of the platform"?
Like seriously that they have just gone all in on higher end hardware for Doom Dark ages was kind of disappointing.
Doom 2016 runs at almost 60fps on a Pentium 4!
The one title that did impress me was ’Humanity’but that uses a heavily modifed Unreal 4 engine, but it all runs very smoothly on limited hardware because of some smart design choices.
That is the literal definition on the first three Google hits I found.
Wikipedia, referencing his original paper:
"The observation is named after Gordon Moore, the co-founder of Fairchild Semiconductor and Intel and former CEO of the latter, who in 1965 noted that the number of components per integrated circuit had been doubling every year,[a] and projected this rate of growth would continue for at least another decade. In 1975, looking forward to the next decade, he revised the forecast to doubling every two years, a compound annual growth rate (CAGR) of 41%. Moore's empirical evidence did not directly imply that the historical trend would continue, nevertheless, his prediction has held since 1975 and has since become known as a law.
Well, yeah. Moore's "law" is subject to the actual laws of physics, and the linearity of advances in transistor density over time is due to humans making it so - human chosen targets for each next generation, but as we come up against the laws of physics and cost of battling them then of course this linear trend will become an asymptote.
Clearly what is driving advances in compute nowadays is not single-chip transistor density but instead multi-chiplet datacenter processors and much higher level multi-chip connectivity such as TPU pods and AI datacenter designs.
I've been reading posts with equally daunting arguments since 1992 on Usenet.
As the article points out, Moore's Law was principally about the number of features. Today, the "Intel ISA" portion of one of their CPU cores is dwarfed by caches and vector & machine learning processing facilities devoted to that same core. Off-chip memory access and other I/O is still a huge factor for why we don't have 10 GHz processors yet. These days we are devoting silicon resources to slowing things down - due to previous performance designs that turned out to have flaws (SPECTRE, Meltdown, etc).
Every obstacle he talks about does seem daunting, and yet I had the same reaction: "haven't we been hearing this for decades?"
The most persuasive argument is the one that he puts in the beginning; at the current rate, we are fast approaching the point where there are no companies left who can afford a new fab.
RE: inversion and speed
Totally fair critique—but HeartMOS™ doesn’t just slow down to save energy. It reengineers inversion to enable both adiabatic flow and GHz-class speed.
Here’s how:
S-Well = Symmetric Inversion
CMOS n/p wells are asymmetrical by default—different V<sub>th</sub>, different μ. That imbalance kills any attempt at clean energy recovery. HeartMOS fixes this with a balanced doping profile (p+ base, tilted retrograde n/p implants) to hit matched threshold voltages and drive currents. That’s symmetric inversion—engineered for resonance.
Fast, Not Fragile
Instead of RC step waves (CMOS square pulses), HeartMOS runs trapezoidal φ/¬φ waveforms at 6GHz, 8-phase pipelined. Think smooth clock rails, not bang-bang toggling. Less dV/dt = less heat = no thermal throttling. It’s not just efficient—it’s sustainable at scale.
RE: inversion and speed
Totally fair critique—but HeartMOS™ doesn’t just slow down to save energy. It reengineers inversion to enable both adiabatic flow and GHz-class speed.
Here’s how:
S-Well = Symmetric Inversion
CMOS n/p wells are asymmetrical by default—different V<sub>th</sub>, different μ. That imbalance kills any attempt at clean energy recovery. HeartMOS fixes this with a balanced doping profile (p+ base, tilted retrograde n/p implants) to hit matched threshold voltages and drive currents. That’s symmetric inversion—engineered for resonance.
Fast, Not Fragile
Instead of RC step waves (CMOS square pulses), HeartMOS runs trapezoidal φ/¬φ waveforms at 6GHz, 8-phase pipelined. Think smooth clock rails, not bang-bang toggling. Less dV/dt = less heat = no thermal throttling. It’s not just efficient—it’s sustainable at scale.
I've always wondered what the classic Moore's-law curve looks like when you take the FLOPs per constant dollar, totaled across the whole R&D/manufacturing/operation process. Sure, you can keep investing more and more into increasingly cutting-edge or power-hungry processes, but at some point it isn't worth the money, and money will generally be the ultimate limiting factor. Not that we'll ever really get these numbers, alas.
> How small or fast or efficient a transistor can be made in a lab is of absolutely no relevance if they can’t be mass-manufactured at a price anyone is willing to pay.
Lowering the cost makes many more things economically viable. (I'm not talking about the fabs here, I'm talking about the customers.) If a new step drops the FLOPS per dollar to the customer, it may create sufficient demand volume to pay for the R&D and the fab, even if those costs are enormous.
> In a transistor, the voltage of the gate lying on top of the channel controls the conductivity of the channel beneath it, either creating an insulator or “depletion region”, or leaving the silicon naturally conductive.
That's... not how this works at all. Eventually the depletion region where the positive or negative charge carriers (for p or n doped silicon) deplete far enough and then at the threshold voltage inversion happens when the opposite sort of charge carrier start to accumulate along the oxide and allow conduction. By surrounding the channel there's less space for a depletion region and so inversion happens at lower voltages, leading to higher performance. Same as people used to do with silicon on oxide.
What's interesting is that with modern GAA and nanosheet transistors at the "1-2nm" scale (really ~6-12nm gate length) is that there are only about 50-100k silicon atoms.
That means that the channels are inherently intrinsic. You can't really have doping when there's statistically less than an atom. There's a nice review from 2023.
The key is that all increases in transistor count are now based on either stacking layers in the silicon (like flash scaling), stacking die with chiplets (we're already at 12-16 die for HBM) or scaling beyond the reticle limit (there's a lot of investment in larger substrates right now). None of these help cost scaling and all have huge yield challenges.
Moore's law really ends when the investors and CFOs decide it does. The generative AI boom has extended that for a while.
from a strict ML perspective, stacking, chiplets, multi-reticle are not going to help.
that is, ML is about periodic shrinks producing squared improvements to device density at iso-area, and thus costs. if you have to change equipment, ML (in this narrow sense) is out the window. if you have to spend linear resources to stack die or chiplets, that's also not ML-compatible (because linear, not 1/(shrink^2))
vertical flash is interesting because it appears to scale surprisingly well - that is, extra layers don't seem to exponentially degrade yield. I'm not sure any of that applies to logic, though.
Yep, only yield improvements and equipment depreciation are really in the cards. Cost per transistor increased from 7nm to 5nm, but we're about at breakeven with 3nm again. If there's not real fab competition at 2nm then prices per transistor likely won't fall. I suspect the same is true at 1nm. It doesn't feel like any of the alternatives (CNT etc) are really able to step up right now either.
My feeling is that algorithmic improvements to LLMs may continue for another order of magnitude or two though. That could continue the scaling for a while.
Of course fixed development costs are now over $1B at 2nm, so if you're not making a large number of die (>1M at $1k/ea), it makes no sense to scale.
> Roughly every five years, the cost to build a factory for making such chips doubles, and the number of companies that can do it halves.
So we may have Apple and NVidia as the only ones that can afford to build a fab. Edit, correction, Microsoft is the current number 2 company by market cap.
You're kind of agreeing with me. According to the article, at some point in the future a new fab will require a joint venture between a tech giant and a chipmaker.
Even one step lower in the chain, ASML sells prototypes of their new photolitho chip makers to the big chip houses years before they develop them. The chip houses understand that the first releases will require ASML engineers to babysit the new tech, patching design issues as they become visible through use.
since then, there have been some adjustments, but it still holds as a prediction of a general trend since as noted in that article:
>One reason for the success of Moore’s prediction is that it became a guide — almost a target — for chip designers.
but as noted:
>The days when we could double the number of transistors on a chip every two years are far behind us. However, Moore’s Law has acted as a pacesetter in a decades-long race to create chips that perform more complicated tasks quicker, especially as our expectations for continual progress continue.
Moore's law assumes photolithography of some form. There are other ways to make chips, and some of them might become viable for mass production soon. Those methods place individual atoms, so smaller features isn't a problem.
I'm waiting somewhat impatiently for AtomicSemi to make some announcements.
When people not super into hardware say Moore's Law, they mean Dennard Scaling.
Which did roughly end in the mid 2000s. That's why we've spent so much time parallelizing in the past 20 years rather than just expecting increases in single threaded perf
There are a lot of people working on all the mentioned problems - and on many many more.
Re garage invention: lithography is probably too big an issue for that. It's important to keep in mind that we're currently producing a lot of transistors with today's tech. Any alternative would have to match that (eg stamping technologies).
An alternative doesn't have to match all capabilities of the current tech. It "only" has to be competitive in one niche, a la The Innovator's Dilemma. Then it can improve and scale from that beach head, like when CMOS went from low-power applications to world domination.
> Rock's law or Moore's second law, named for Arthur Rock or Gordon Moore, says that the cost of a semiconductor chip fabrication plant doubles every four years.[1]
I think you could very easily give a cap that hinges on our current understanding of basic physical limitations, and it would arrive surprisingly soon.
That's the thing about Moore's law - it has assumed from the beginning that our 'current understanding of basic physical limitations' is incomplete, and been proven correct on that front many times over.
Our understanding of basic physical limits seems reasonably good and hasn't changed for a couple of generations. Our understanding of engineering limitations on the other hand is not so good and subject to frequent change.
As I understand it Moore's Law doesn't address any sort of fundamental physical limitations other than perhaps an absolutely limit in terms of some fundamental limit on the smallness of an object, it's just an observation of the doubling of transistor density over a consistent period of time.
It seems more like an economical or social observation than a physical one to me.
I was using OP's terminology, pointing out that people (including Moore himself) have been warning of an imminent cap on Moore's law since at least 1975. Getting into fine detail of what constitutes a 'fundamental physical limit' as opposed to an engineering challenge wasn't really the point, though I personally believe that our understanding of physical limits will develop further.
We don't know what we don't know - there's always the potential of radical technology coming from an upending of things which were 'established' for decades or centuries previously; that's just the nature of science.
In contexts like these, “universe” means the observable universe, which is finite in size. Also, creating universes (in the usual models) conserves energy, so you don’t actually gain anything by that.
> In contexts like these, “universe” means the observable universe, which is finite in size. Also, creating universes (in the usual models) conserves energy, so you don’t actually gain anything by that.
"In the usual models" - that's the whole point. For the entire 60 year history of Moore's Law, the "usual models" have stated that it would come to an end in about ten years.
Roughly every two years, the density of transistors that can be fit onto a silicon chip doubles.
No. Moore's law is not about density. It's just about the number of transistors on a chip. Yes, density increases but so does die size. Anyways, in Moore's own word:
The complexity for minimum component costs has increased at a rate of roughly a factor of two per year.
> Gordon Moore always emphasized that his “law” was fundamentally rooted in economics, not physics.
Effectively, it was always more of a "marketing law" than an engineering one. Semiconductor chips only had 18-36 months to reap big profits, so Intel tried to stay ahead of that curve.
> Effectively, it was always more of a "marketing law" than an engineering one.
Yes, but the engineering still has to be feasible. If Henry Ford had coined a similar law about eg the power of car engines, that might have been interesting marketing, but it doesn't mean Ford and the car industry could have sustained that for as long as Moore's law held.
That's a common misconception. Moore's 1965 paper was about economics but when the phrase "Moore's Law was coined in 1975 it was referring to Dennard Scaling as a whole.
ugh, "scraping the bottom of the barrel" makes me think this is motivated.
especially after describing all the other desperate scaling techniques in the past - and that history argues for the need (almost manifest destiny in the ML sense) to do these things.
One point the article doesn't make explicit: The exponential decrease in transistor size (≈Moore's law) was historically only valuable because it also caused
1) exponentially decreasing cost per transistor and
2) exponentially decreasing power consumption per transistor.
However, in recent years 1) has generally weakened. At some point, the price per transistor will no longer decrease for smaller process nodes, and even start to increase. Then making smaller transistors could only be justified for power constrained chips that benefit from 2). But even this has only limited value.
So at some point, producing chips with smaller transistors will no longer make economic sense, even if transistor size could technically still be decreased at a similar rate as in the past.
the law delivered enough headroom that systems moved on. once compute got cheap to rent and scale ,there was less pressure to push frequency or density every cycle. so focus shifted. the gains kept coming ,just not in the same shape.
It tells us more about market and ppl's hunger for apps/cabilities than fabrication and physics, although fabrication quality is indirectly related through cost.
Current GPUs have a comparable number of transistors (92.2 billion in the current NVidia Blackwell according to https://chipsandcheese.com/p/blackwell-nvidias-massive-gpu) to the number of neurons in human brains (about 90 billion according to Wikipedia). Brains consume less energy and do more, though transistors beat them on density. This suggests there are alternative pathways to performing computation that will scale better.
Let me restate. The article is musing about medium terms difficulties on the current pathway for producing computation. I'm musing that perhaps we are on the wrong pathway for producing computation.
That's the best we have, and not for lack of trying. There are attempts at analog inference (like Mythic) but it doesn't look as they are going anywhere yet.
> Another possibility that has long been on my personal list of “future articles to write” is that the future of computing may look more like used cars. If there is little meaningful difference between a chip manufactured in 2035 and a chip from 2065, then buying a still-functional 30-year-old computer may be a much better deal than it is today. If there is less of a need to buy a new computer every few years, then investing a larger amount upfront may make sense – buying a $10,000 computer rather than a $1,000 computer, and just keeping it for much longer or reselling it later for an upgraded model.
This seems improbable.
50-year-old technology works because 50 years ago, transistors were micron-scale.
Nanometer-scale nodes wear out much more quickly. Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
One of my concerns is we're reaching a point where the loss of a fab due to a crisis -- war, natural disaster, etc. -- may cause systemic collapse. You can plot lifespan of chips versus time to bring a new fab online. Those lines are just around the crossing point; modern electronics would start to fail before we could produce more.
> Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
That statement absolutely needs a source. Is "usage" 100% load 24/7? What is the failure rate after 7 years? Are the failures unrepairable, i.e. not just a broken fan?
Yeah, chip aging is at max temperature, max current, and worst process corner. And it's nonlinear so running at <10% duty cycle could reduce aging to almost nothing.
Has chip aging finally surpassed thermal cycling as the primary cause of component failure in datacenters?
I don't know but I would guess not. Solder is really weak compared to silicon.
Every now and then, I get a heartfelt chuckle from HN.
By 'Modern' they must mean latest generation, so we'll have to wait and see. I was imagining not using an RTX 5090 for 7 years and find it doesn't work, or one used 24x7 for 3 years then failing.
Electromigration and device aging are huge issues. I can't imagine a modern GPU having a lifetime longer than 3 years at 100C temperature.
Though, it can be solved with redundancy at the cost of performance.
I’ve never heard of this and I was an Ethereum miner. We pushed the cards as hard as they would go and they seemed fine after. As long as the fan was still going they were good.
So Intel used to claim a 100,000+ hour life time on their chips. They didnt actually test them to this because that is 11.4 years. But it was basically saying, these things will last at full speed way beyond any reasonable life time. Many chip could probably go way beyond that.
I think it was about 15 years back they stopped saying that. Once we passed the 28nm mark it started to become apparent that they couldnt really state that.
It makes sense, as parts get smaller they will get more fragile from general usage.
With your GPUs yeah they are probably still fine but they could already be half way through their life time, you wouldnt know it until failure point. Add in the silicon lotto and it gets more complicated.
One thing to realize is the lifetime is a statistical thing.
I design chips in modern tech nodes (currently using 2nm). What we get feom the fab is a statistical model of device failure modes. Aging is one of them. When transistors gradually age they get slower sue to increased threshold voltage. This eventually causes failure at a point where timing is tight. When will it happen varies greatly sue to initial conditions, exact conditions the chip was in(temp, vdd, number of on-off cycles, even the workload). After an agong failure the chip will still work if the clock freq is reduced. There are aging monitors on-chip sometimes which try to catch it early and scale down the clock.
There are catastrophic failures too, like gate insulator breakdown, electromigration or mechanical failures of IO interconnect. The last one is orders of magnitude more likely than anything else these days.
For mining, If a GPU was failing in such a way that it was giving completely wrong output for functions during mining, that'd only be visible as a lower success hash-rate which you might not even notice unless you did periodic testing of known-target hashes.
For graphics, the same defect could be severe enough to completely render the GPU useless.
Just look at warranties, gotta go to Quadro series for industrial warranty lengths.
> Nanometer-scale nodes wear out much more quickly. Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
I recently bought a new MacBook, my previous one having lasted me for over 10 years. The big thing that pushed me to finally upgrade wasn’t hardware (which as far as I could tell had no major issues), it was the fact that it couldn’t run latest macOS, and software support for the old version it could run was increasingly going away.
The battery and keyboard had been replaced, but (AFAIK) the logic board was still the original
> it couldn’t run latest macOS, and software support for the old version it could run was increasingly going away.
which is very annoying, as none of the newer OS versions has anything that warrants dumping hardware to buy brand new to run them with! With the exception of security upgrades, which i find dubious for a company to stop creating (as they would need to do so for their newer OS versions just as well, so the cost of maintaining security patches ought to not be much, if at all), it is definitely more likely to be a dark-pattern to force hardware upgrades.
That's not just a dark pattern, it's the logical conclusion to Apple's entire business model. It's what you get for relying on the proprietary OS supplied by a hardware manufacturer. It's why Asahi Linux is so important.
I'm not sure I agree. Open source software also regularly drops support for old hardware and OSes.
"regularly" is doing a lot of work here. When Linux drops hardware support, we are talking about ancient hardware. An example of a regular drop: Linux 6.15 just a month ago dropped support for 486 (from 1989)!
Pretty much, you can still get modern distros that support 32bit PowerPC.
Open source software drops hardware support only when there are nobody left who volunteers to support that hardware. When does this happen? It happens when there are not enough users left of that hardware.
As long as there are enough users of some hardware, free software will support it, because the users of that hardware want it to.
Is "regularly" every 2-4 years, or longer? What are your options? With Apple you have none. It's really not a comparable situation.
And then he still couldn’t use the third party software he says he depends on…
Depending on how much has changed in the interval, backporting security fixes can be completely trivial, very difficult, or anywhere in between. There may not even be a fix to backport, as not all vulnerabilities are still present in the latest release.
You mean besides the fact that they completely transitioned to a new processors and some of the new features use hardware that is only available on their ARM chips?
Also he said that software from third parties also don’t support the older OS so even if Apple did provide security updates, he would still be in the same place.
I've got 3 Macbooks from 2008, 2012, and 2013. Apple dropped MacOS support years ago. They all run the latest Fedora Linux version with no problems.
The screen on the MacBookPro10,2 is 2560x1600 which is still higher than a lot of brand new laptops. The latest version it will run is 10.15 from 2019. I know Apple switched to ARM but most people don't need a new faster computer. I stopped buying Apple computers because I want my computer supported more than 6 years.
I do have 3 newer computers but these old Macbooks are kept at various relative's houses for when I visit and wnat my own Linux machine. They have no problems running a web browser and watching videos so why replace them?
> Modern GPUs have a rated lifespan in the 3-7 year range, depending on usage.
I seriously doubt this is true. The venerable GTX 1060 came out 9 years ago, and still sees fairly widespread use based on the Steam hardware survey. According to you, many (most?) of those cards should have given out years ago.
This is just untrue, and you’ve provided no citation, either.
The silicon gates in GPUs just don’t wear out like that, not at that timescale. The only thing that sort of does is SSDs (and that’s a write limit, which has existed for decades, not a new thing).
Electromigration tends to get worse with small sizes but also higher voltage and temperatures. I could see a GPU wearing out that quickly if it were overclocked enough, but stock consumer GPUs will last much longer than that.
electromigration is real, but is it relevant?
since electromigration is basically a matter of long, high-current interconnect, I guess I have been assuming it's merely designed around. By, for instance, having hundreds of power and ground pins, implying quite a robust on-chip distribution mesh, rather than a few high-current runs.
Wouldnt it depend on work loads? My GPU that kicks into high gear for maybe 2-3 hours a week will probably do decades of use before chip degradation kicks in. The power capacitors will give out long before the silicon does.
But it someone is running an LLM 24 hours a day, might not go for as long.
We are flying blind, both on those claiming short life span and those who are not.
you're thinking of what, electromigration?
what is the age-related failure mode you're referring to?
or are you merely referring to warranty period? (which has more to do with support costs, like firmware - not expected failures.)
> One of my concerns is we're reaching a point where the loss of a fab due to a crisis -- war, natural disaster, etc. -- may cause systemic collapse.
This is absolutely ridiculous. Even if Taiwan sank today we really don't need those fabs for anything critical. i strongly suspect we could operate the entire supply chain actually necessary for human life with just z80s or some equivalent.
More parallelism. Less clock. More l1 cache per CPU and less disk stalls. Are there plenty of tricks in the sea, when clockspeed goes flat?
Dual ported TCAM memory isn't getting faster and we've got to 1,000,000 prefixes in the internet and ipv6 are 4 times bigger. Memory speed is a real issue.
Moore's law isn't about clockspeed.
There has been a relationship between die area and speed for much of time. That stopped being true. Perhaps it was only a side effect of shorter paths on chip.
You might be talking about https://en.wikipedia.org/wiki/Dennard_scaling
The question is at what cost? Cache hasnt scaled down for a long while now, it is part of why it takes up so much die space nowadays.
Dont mistake technical viability for economic viability.
What does economics in die area say about using die for highly repeatable simple memory forms, vs complex instruction/ALU? I would think there is a balance point between more cores, and faster cores with more cache and more lanes between them.
We're on 8 as a mixture of fast and slow, hot and cool, on almost all personal devices now. I routinely buy Intel instruction racks with 16 per CPU construct, and a TB of memory. At one remove, chips seem to be fulfilling my mission but I am not in GPU/TPU land. Although I note, FPGA/GPU/TPU seems like where a lot of the mental energy goes these days.
VLSI is a black art. Quantum effects are close enough to magic I feel confident saying few people today understand VLSI. I know I don't, I barely understand virtual memory. I ache for the days when a zero pointer was a valid address on a pdp11
I'm wondering if cache is due for a one time doubling of line length. Apple did it, but they seem to be unique in that.
I think its possible. When 128 bits came along, I wondered if the very long instruction word people were going to come back into the room. If we imagine mapping addressing into the world, how about making some rules about top bits addressing other nodes in the DC and see how "shared memory" works between distinct racks?
Moore’s law has been unsustainable for 20 years, I remember Pentium 4’s with 4ghz. But that hasn’t seemed to matter in terms of real day to day performance improvements. This article makes some great points about the scaling cost and reduced market opportunity for there to be more than 2 or 3 makers in the market, but that’s a trend we’ve seen in every market in the world, to be honest I’m surprised it took this long to get there.
As interesting as this breakdown of the current state of things is, it doesn’t tell us much we didn’t know or predict much about the future, and that’s the thing I most wanted to hear from an expert article on the subject, even if we can take it with a large pinch of salt.
The corollary is Wirth's Law:
>software is getting slower more rapidly than hardware is becoming faster.
>Wirth attributed the saying to Martin Reiser, who in the preface to his book on the Oberon System wrote: "The hope is that the progress in hardware will cure all software ills. However, a critical observer may observe that software manages to outgrow hardware in size and sluggishness."
I wish that there would be more instances of developments like to Mac OS X 10.6, where rather than new features, the software was simply optimized for a given CPU architecture, and the focus was on improving performance.
This doesn’t hold with my experience, I remember how slow my old machines used to be - especially the ones with spinning hard disks. Not to mention early smartphones… I’ve had computers in my youth I used to turn on then go and make a coffee. Now it’s practically instant.
> Now it’s practically instant.
No it's not :)
Phones do not shut down. Powering off and back on still takes forever.
As for desktops, Windows at least keeps starting stuff in the background long after your UI gets displayed.
Yeeeah dude you have some seriously rose tinted glasses on.
My windows machine can go from completely powered off to a big application like a web browser fully open and operational on maybe 10-15 seconds, and yes, fully powered off. No way I’m making coffee in that time. Sure, windows might run stuff in the background, but it’s so fast that stuff doesn’t get in the way of me using the thing at all.
As for phones, “forever” is a silly thing to say but even if I granted you that one off boot time, you have somewhat moved the goalposts as you talked about the act of using software in day to day usage. On my laptop I can genuinely open 10 apps, close 10 apps, and open 10 apps again all within 10 seconds, probably faster if I wasn’t constrained by literal click actions. My computers growing up would grind to a halt if you tried to run like 3 things at once!
There are always instances of inefficient software, but there was back in the day too, I remember generating a new map on civilisation 1 on my 286 and it taking 10 hours. There is no way you can possibly say that on average software is slower today than it used to be, that statement in general is just completely false.
> I remember generating a new map on civilisation 1 on my 286 and it taking 10 hours
Humm. If my glasses are rose tinted, yours are black tinted. I've played quite a lot of civ 1, even on an xt with 8088, and i'm pretty sure map generation did not take 10 hours.
Perhaps if you hacked yours to use larger maps...
I'm also pretty sure civ 1 got to the menu on my 8088 faster than Death Stranding 2 on my ps5. Although that might be my rose tinted glasses indeed. Civ 1 also didn't have a 29 Gb day one patch :)
The improvements you describe come down to one thing: solid state drives. If we had them in the 90s, your software would've been starting up just as fast (or even faster) then. Meanwhile, once the software did start, it ran faster and with fewer resources than software does today. It's not rose colored glasses: software today truly is worse on average than software 20 years ago.
The horrors of booting Windows 2000 on a Pentium 1.
But there is something to be said about prior optimizations in limited resources. A few months back a friend was restoring a G4 iMac and I was astounded at how snappy the whole thing felt once the OS had finally loaded.
We forget how snappy iTunes was while using only 30MB of RAM.
If you can combine the best of boths worlds, there could be something special there.
Modern OS and software are very bloated but it is astounding how fast the hardware crunches through all that. But it could be better.
And yet https://danluu.com/input-lag/
Software optimized for a particular processor is today's tech. In the future we may see CPUs that alter themselves to better serve the needs of software. I can see a day where a "cpu" was actually a bunch of different processors connected via a FPGA so that the software could reconfigure the CPU on the fly.
Why the down votes here? This is an interesting idea.
The great thing about AI, it is finally a killer feature that is enjoyed and useful by every user worldwide. And the tech industry finally have an excuse to up sell 16GB as baseline, and perhaps even try to push 24 or 32GB Memory, along with GPU / NPU and CPU upgrade.
For users, a few hundred dollar extra ( on top of the original purchase ) is a such a small number compared to the productivity gain over the usage span of the computer.
AI alone not only increased the server hardware requirement but also user client requirement. It is basically the question everyone has been asking, what is after Smartphone? And to a degree it is AI. ( or LLM )
This will easily push the whole Semi-Conductor Industry forward all the way to 2032 ~ 2035. We will be at 8A or 6A by then.
PCIe 7? possibly PCIe 8? WiFi 9 which is a fixed version of WiFi 8. There are so many great Hardware improvement coming out all because of the demand of greater computing usage.
Software side has been rather boring TBH. I really like the phase Allan Kay uses to describe modern days software are "reinventing the flat tire".
> The great thing about AI, it is finally a killer feature that is enjoyed and useful by every user worldwide.
We must live in very different worlds. Most people I know absolutely hate it and are immediately put off by anything that tries to force it in.
I am not AI but it needs to be used purposefully rather than just another check box to the feature set.
Productivity gains for some, needless slop for others.
Anecdotes don't mean much. For example, most people I know use it. It's just a very normal technology now.
For something more substantial than an anecdote, 92% of UK students use AI: https://studenttimes.org/news/uk-students-turn-to-ai-in-reco...
I think the future of compute will look much like today!
Given the power and ubiquity of smart phones, most people don't need any other computer in their personal life. What can be done locally on a smart phone seems like it will be more constrained by battery life and physical size than anything else, and there will continue to be a mix of things that can run on-device and other more compute-hungry functions run in the cloud. I don't see smartphones being replaced or augmented by other devices like smart glasses - people want one device that does it all, and not one they wear on their face.
The same is somewhat true for business use too, especially if compute-heavy AI use becomes more widespread - some functions local and the heavy work done in AI-ready datacenters. I'm mildly surprised that there hasn't already been a greater shift away from local compute to things like Chromebooks (ubiquitous in schools) since it has so many benefits (cost, ease of management, reduced need to upgrade), but maybe it will still come if the need to rely on datacenter compute increases.
Even if we imagine futuristic power-sipping neuromorphic chips, I don't see that changing things very much other that increasing the scope of what can be done locally on a power budget.
> Given the power and ubiquity of smart phones, most people don't need any other computer in their personal life.
Maybe, but they need other form factors, too. Like a big screen and a keyboard.
And once you have those, you might as well through some compute in there: getting smartphone-like performance out of desktop-like components (or even laptop components) is pretty cheap, so might as well throw them in.
Chromebooks are fairly capable 'thin' clients. Their specs aren't that much worse than other laptops.
However all that being said, I mostly agree that at least with current style AI we will likely see a lot more data centre compute in the future. At least as long as hardware costs are a significant fraction of total costs: the 'cloud' can multiplex multiple users over expensive hardware, your local computer is mostly idle and waiting for your input. (And it has to be, if you want low latency.)
However, if costs shift so that electricity and not hardware is the main cost, we might see a return to local compute. Maybe.
> Maybe, but they need other form factors, too. Like a big screen and a keyboard.
It seems generational habits are changing. TV is dead, and younger generation seems ok watching things on small screens, even smartphone. Can always sync phone to parent's/hotel big screen TV if available.
That's Dennard scaling, Moore's law is not about clocks but components on an integrated circuit.
God of war 2 was made on 300mhz cpu and 32mb of ram.
We haven't been bound by moore's law because we just waste computing power because programmers are expensive. No one is trying to optimize nowadays except in very niche places. And when push comes to shove we just start wasting slightly less. Like adding a JIT to a script language 15 years too late.
As someone who has met people whose entire job is to optimise games for consoles, this isn’t true. It’s ok to want to get the most out of hardware in terms of features too.
Can you point to a couple of games in the last few years that would make one exclaim "what kind of forbidden magic they did so they squeeze so much performance out of the platform"?
Doom... 8 years ago...
Like seriously that they have just gone all in on higher end hardware for Doom Dark ages was kind of disappointing.
Doom 2016 runs at almost 60fps on a Pentium 4!
The one title that did impress me was ’Humanity’but that uses a heavily modifed Unreal 4 engine, but it all runs very smoothly on limited hardware because of some smart design choices.
> Roughly every two years, the density of transistors that can be fit onto a silicon chip doubles. This is Moore’s Law.
that... isn't the moore law, it is about count / complexity, not density. and larger chips are a valid way to fullfill it.
https://hasler.ece.gatech.edu/Published_papers/Technology_ov...
https://www.eng.auburn.edu/~agrawvd/COURSE/E7770_Spr07/READ/...
That is the literal definition on the first three Google hits I found.
Wikipedia, referencing his original paper:
"The observation is named after Gordon Moore, the co-founder of Fairchild Semiconductor and Intel and former CEO of the latter, who in 1965 noted that the number of components per integrated circuit had been doubling every year,[a] and projected this rate of growth would continue for at least another decade. In 1975, looking forward to the next decade, he revised the forecast to doubling every two years, a compound annual growth rate (CAGR) of 41%. Moore's empirical evidence did not directly imply that the historical trend would continue, nevertheless, his prediction has held since 1975 and has since become known as a law.
Why you quoting Wikipedia to me?
I literally linked the paper. The paper doesn't mention density.
"First three website you found on Google"'s opinion don't matter for a lot when we have the original formulation, right there, archived
And Wikipedia passage you quoted doesn't mention density either, only count.
I really, truly don't understand where your comment is coming from and going to.
https://news.ycombinator.com/item?id=44412083
even if it were, that isn't about transistor density, but power density, which is not the same as
> the density of transistors that can be fit onto a silicon chip doubles
the whole article takes off from a flawed and fantasious misinterpretation and argue against that self created windmill
Well, yeah. Moore's "law" is subject to the actual laws of physics, and the linearity of advances in transistor density over time is due to humans making it so - human chosen targets for each next generation, but as we come up against the laws of physics and cost of battling them then of course this linear trend will become an asymptote.
Clearly what is driving advances in compute nowadays is not single-chip transistor density but instead multi-chiplet datacenter processors and much higher level multi-chip connectivity such as TPU pods and AI datacenter designs.
I've been reading posts with equally daunting arguments since 1992 on Usenet.
As the article points out, Moore's Law was principally about the number of features. Today, the "Intel ISA" portion of one of their CPU cores is dwarfed by caches and vector & machine learning processing facilities devoted to that same core. Off-chip memory access and other I/O is still a huge factor for why we don't have 10 GHz processors yet. These days we are devoting silicon resources to slowing things down - due to previous performance designs that turned out to have flaws (SPECTRE, Meltdown, etc).
The features per nm just keep on a-coming.
Every obstacle he talks about does seem daunting, and yet I had the same reaction: "haven't we been hearing this for decades?"
The most persuasive argument is the one that he puts in the beginning; at the current rate, we are fast approaching the point where there are no companies left who can afford a new fab.
Yeah I feel like this article should have (1990)
> is still a huge factor for why we don't have 10 GHz processors yet
You're on air cooled silicon. You were never going past 4GHz in this configuration. This was an understood limit well before we hit it.
RE: inversion and speed Totally fair critique—but HeartMOS™ doesn’t just slow down to save energy. It reengineers inversion to enable both adiabatic flow and GHz-class speed.
Here’s how:
S-Well = Symmetric Inversion CMOS n/p wells are asymmetrical by default—different V<sub>th</sub>, different μ. That imbalance kills any attempt at clean energy recovery. HeartMOS fixes this with a balanced doping profile (p+ base, tilted retrograde n/p implants) to hit matched threshold voltages and drive currents. That’s symmetric inversion—engineered for resonance.
Fast, Not Fragile Instead of RC step waves (CMOS square pulses), HeartMOS runs trapezoidal φ/¬φ waveforms at 6GHz, 8-phase pipelined. Think smooth clock rails, not bang-bang toggling. Less dV/dt = less heat = no thermal throttling. It’s not just efficient—it’s sustainable at scale.
Bottom line: Adiabatic ≠ slow. Asymmetric inversion = slow. Symmetric = harmonic = fast.
Patent Pending – “HeartMOS™ — Symmetric Semiconductor Device with Co-Optimized Doping for Reversible Computing” https://sites.tufts.edu/tcal/publications/hotgauge/
RE: inversion and speed Totally fair critique—but HeartMOS™ doesn’t just slow down to save energy. It reengineers inversion to enable both adiabatic flow and GHz-class speed.
Here’s how:
S-Well = Symmetric Inversion CMOS n/p wells are asymmetrical by default—different V<sub>th</sub>, different μ. That imbalance kills any attempt at clean energy recovery. HeartMOS fixes this with a balanced doping profile (p+ base, tilted retrograde n/p implants) to hit matched threshold voltages and drive currents. That’s symmetric inversion—engineered for resonance.
Fast, Not Fragile Instead of RC step waves (CMOS square pulses), HeartMOS runs trapezoidal φ/¬φ waveforms at 6GHz, 8-phase pipelined. Think smooth clock rails, not bang-bang toggling. Less dV/dt = less heat = no thermal throttling. It’s not just efficient—it’s sustainable at scale.
Bottom line: Adiabatic ≠ slow. Asymmetric inversion = slow. Symmetric = harmonic = fast.
I've always wondered what the classic Moore's-law curve looks like when you take the FLOPs per constant dollar, totaled across the whole R&D/manufacturing/operation process. Sure, you can keep investing more and more into increasingly cutting-edge or power-hungry processes, but at some point it isn't worth the money, and money will generally be the ultimate limiting factor. Not that we'll ever really get these numbers, alas.
The author makes a related point:
> How small or fast or efficient a transistor can be made in a lab is of absolutely no relevance if they can’t be mass-manufactured at a price anyone is willing to pay.
Lowering the cost makes many more things economically viable. (I'm not talking about the fabs here, I'm talking about the customers.) If a new step drops the FLOPS per dollar to the customer, it may create sufficient demand volume to pay for the R&D and the fab, even if those costs are enormous.
> In a transistor, the voltage of the gate lying on top of the channel controls the conductivity of the channel beneath it, either creating an insulator or “depletion region”, or leaving the silicon naturally conductive.
That's... not how this works at all. Eventually the depletion region where the positive or negative charge carriers (for p or n doped silicon) deplete far enough and then at the threshold voltage inversion happens when the opposite sort of charge carrier start to accumulate along the oxide and allow conduction. By surrounding the channel there's less space for a depletion region and so inversion happens at lower voltages, leading to higher performance. Same as people used to do with silicon on oxide.
The Wikipedia article has nice diagrams:
https://en.wikipedia.org/wiki/MOSFET
What's interesting is that with modern GAA and nanosheet transistors at the "1-2nm" scale (really ~6-12nm gate length) is that there are only about 50-100k silicon atoms.
That means that the channels are inherently intrinsic. You can't really have doping when there's statistically less than an atom. There's a nice review from 2023.
https://semiengineering.com/what-designers-need-to-know-abou...
https://www.semiconductor-digest.com/the-shape-of-tomorrows-...
This is more recent with pretty pictures.
The key is that all increases in transistor count are now based on either stacking layers in the silicon (like flash scaling), stacking die with chiplets (we're already at 12-16 die for HBM) or scaling beyond the reticle limit (there's a lot of investment in larger substrates right now). None of these help cost scaling and all have huge yield challenges.
Moore's law really ends when the investors and CFOs decide it does. The generative AI boom has extended that for a while.
from a strict ML perspective, stacking, chiplets, multi-reticle are not going to help.
that is, ML is about periodic shrinks producing squared improvements to device density at iso-area, and thus costs. if you have to change equipment, ML (in this narrow sense) is out the window. if you have to spend linear resources to stack die or chiplets, that's also not ML-compatible (because linear, not 1/(shrink^2))
vertical flash is interesting because it appears to scale surprisingly well - that is, extra layers don't seem to exponentially degrade yield. I'm not sure any of that applies to logic, though.
Yep, only yield improvements and equipment depreciation are really in the cards. Cost per transistor increased from 7nm to 5nm, but we're about at breakeven with 3nm again. If there's not real fab competition at 2nm then prices per transistor likely won't fall. I suspect the same is true at 1nm. It doesn't feel like any of the alternatives (CNT etc) are really able to step up right now either.
https://www.chipstrat.com/p/what-happens-if-tsmc-controls-it
My feeling is that algorithmic improvements to LLMs may continue for another order of magnitude or two though. That could continue the scaling for a while.
Of course fixed development costs are now over $1B at 2nm, so if you're not making a large number of die (>1M at $1k/ea), it makes no sense to scale.
> Roughly every five years, the cost to build a factory for making such chips doubles, and the number of companies that can do it halves.
So we may have Apple and NVidia as the only ones that can afford to build a fab. Edit, correction, Microsoft is the current number 2 company by market cap.
None of these companies know how to build a fab. The most they could do is invest in TSMC building one as a Joint Venture.
You're kind of agreeing with me. According to the article, at some point in the future a new fab will require a joint venture between a tech giant and a chipmaker.
It's already the case.
Even one step lower in the chain, ASML sells prototypes of their new photolitho chip makers to the big chip houses years before they develop them. The chip houses understand that the first releases will require ASML engineers to babysit the new tech, patching design issues as they become visible through use.
That is a de facto joint venture.
They can't afford to tank their margins like that, investors would be rather unhappy.
Has "Moore's Law" been consistent since it reared its head, or has it been constantly tweaked to suit the narrative of it still being correct?
It was good up until 1975:
https://www.livescience.com/technology/electronics/what-is-m...
since then, there have been some adjustments, but it still holds as a prediction of a general trend since as noted in that article:
>One reason for the success of Moore’s prediction is that it became a guide — almost a target — for chip designers.
but as noted:
>The days when we could double the number of transistors on a chip every two years are far behind us. However, Moore’s Law has acted as a pacesetter in a decades-long race to create chips that perform more complicated tasks quicker, especially as our expectations for continual progress continue.
Moore's law assumes photolithography of some form. There are other ways to make chips, and some of them might become viable for mass production soon. Those methods place individual atoms, so smaller features isn't a problem.
I'm waiting somewhat impatiently for AtomicSemi to make some announcements.
The last sentence should be first, to put things in context:
> We’re entering the post-Moore era, I’m busy designing chips (and maybe a fab) for this new world. I’d be happy to talk to investors.
We were told in uni in the early 2000s that post-Moore era was just few years from then.
When people not super into hardware say Moore's Law, they mean Dennard Scaling.
Which did roughly end in the mid 2000s. That's why we've spent so much time parallelizing in the past 20 years rather than just expecting increases in single threaded perf
There are a lot of people working on all the mentioned problems - and on many many more.
Re garage invention: lithography is probably too big an issue for that. It's important to keep in mind that we're currently producing a lot of transistors with today's tech. Any alternative would have to match that (eg stamping technologies).
(I work on lithography optics)
An alternative doesn't have to match all capabilities of the current tech. It "only" has to be competitive in one niche, a la The Innovator's Dilemma. Then it can improve and scale from that beach head, like when CMOS went from low-power applications to world domination.
Is there also a law for how much more difficult it becomes to sustain Moore's law?
Ultimately, there's a cap. For as far as I know, the universe is finite.
Landuaer's principle govern's how efficient computation can be, but we might have to transition to something other than transistors to hit that limit.
https://en.wikipedia.org/wiki/Landauer%27s_principle
> Is there also a law for how much more difficult it becomes to sustain Moore's law?
https://en.wikipedia.org/wiki/Moore%27s_second_law
> Rock's law or Moore's second law, named for Arthur Rock or Gordon Moore, says that the cost of a semiconductor chip fabrication plant doubles every four years.[1]
> as far as I know, the universe is finite.
I don't think we know that. We don't even know how big the universe really is - we can only see so far. All we have is a best guess.
There may also be a multiverse out there (or right beside us).
And, creating universes might be a thing.
... I don't expect Moore's law to hold for ever either, but I don't believe in creating unnecessary caps.
I think you could very easily give a cap that hinges on our current understanding of basic physical limitations, and it would arrive surprisingly soon.
That's the thing about Moore's law - it has assumed from the beginning that our 'current understanding of basic physical limitations' is incomplete, and been proven correct on that front many times over.
Our understanding of basic physical limits seems reasonably good and hasn't changed for a couple of generations. Our understanding of engineering limitations on the other hand is not so good and subject to frequent change.
I'm not sure I follow. can you elaborate on that?
As I understand it Moore's Law doesn't address any sort of fundamental physical limitations other than perhaps an absolutely limit in terms of some fundamental limit on the smallness of an object, it's just an observation of the doubling of transistor density over a consistent period of time.
It seems more like an economical or social observation than a physical one to me.
I was using OP's terminology, pointing out that people (including Moore himself) have been warning of an imminent cap on Moore's law since at least 1975. Getting into fine detail of what constitutes a 'fundamental physical limit' as opposed to an engineering challenge wasn't really the point, though I personally believe that our understanding of physical limits will develop further.
We don't know what we don't know - there's always the potential of radical technology coming from an upending of things which were 'established' for decades or centuries previously; that's just the nature of science.
In contexts like these, “universe” means the observable universe, which is finite in size. Also, creating universes (in the usual models) conserves energy, so you don’t actually gain anything by that.
> In contexts like these, “universe” means the observable universe, which is finite in size. Also, creating universes (in the usual models) conserves energy, so you don’t actually gain anything by that.
"In the usual models" - that's the whole point. For the entire 60 year history of Moore's Law, the "usual models" have stated that it would come to an end in about ten years.
if you want to be a "universe-alist", the mass of all those transistors would collapse into a black hole before you ran out of space.
> Gordon Moore always emphasized that his “law” was fundamentally rooted in economics, not physics.
Effectively, it was always more of a "marketing law" than an engineering one. Semiconductor chips only had 18-36 months to reap big profits, so Intel tried to stay ahead of that curve.
> Effectively, it was always more of a "marketing law" than an engineering one.
Yes, but the engineering still has to be feasible. If Henry Ford had coined a similar law about eg the power of car engines, that might have been interesting marketing, but it doesn't mean Ford and the car industry could have sustained that for as long as Moore's law held.
That's a common misconception. Moore's 1965 paper was about economics but when the phrase "Moore's Law was coined in 1975 it was referring to Dennard Scaling as a whole.
https://hopefullyintersting.blogspot.com/2019/03/what-is-moo...
factor in power usage reduction, and it still works
ugh, "scraping the bottom of the barrel" makes me think this is motivated.
especially after describing all the other desperate scaling techniques in the past - and that history argues for the need (almost manifest destiny in the ML sense) to do these things.
One point the article doesn't make explicit: The exponential decrease in transistor size (≈Moore's law) was historically only valuable because it also caused
1) exponentially decreasing cost per transistor and
2) exponentially decreasing power consumption per transistor.
However, in recent years 1) has generally weakened. At some point, the price per transistor will no longer decrease for smaller process nodes, and even start to increase. Then making smaller transistors could only be justified for power constrained chips that benefit from 2). But even this has only limited value.
So at some point, producing chips with smaller transistors will no longer make economic sense, even if transistor size could technically still be decreased at a similar rate as in the past.
the law delivered enough headroom that systems moved on. once compute got cheap to rent and scale ,there was less pressure to push frequency or density every cycle. so focus shifted. the gains kept coming ,just not in the same shape.
> This makes a valuable asset that would otherwise be prohibitively expensive affordable to a much larger portion of the population.
They just have to deal with less efficiency and higher maintenance costs. You have not discovered a free lunch, sorry.
So tired of people getting this wrong. Moore's law is about NUMBER of transistors, NOT DENSITY!
https://en.m.wikipedia.org/wiki/Moore%27s_law
It tells us more about market and ppl's hunger for apps/cabilities than fabrication and physics, although fabrication quality is indirectly related through cost.
Thank you for the nice downvotes when I posted correct information including a source.
I'm not upvoting due to your aggro tone & capslock, if that helps.
[dead]
Current GPUs have a comparable number of transistors (92.2 billion in the current NVidia Blackwell according to https://chipsandcheese.com/p/blackwell-nvidias-massive-gpu) to the number of neurons in human brains (about 90 billion according to Wikipedia). Brains consume less energy and do more, though transistors beat them on density. This suggests there are alternative pathways to performing computation that will scale better.
It takes many transistors to replicate a single neuron, they work very differently in terms of speed, there is no direct comparison.
Let me restate. The article is musing about medium terms difficulties on the current pathway for producing computation. I'm musing that perhaps we are on the wrong pathway for producing computation.
This only applies to functions of the brain which we wish to replicate on computers, not to those computers already outperform us on.
Agreed.
That's the best we have, and not for lack of trying. There are attempts at analog inference (like Mythic) but it doesn't look as they are going anywhere yet.