I thought Moore's Law effectively ended over a decade ago when we switched to multi-core chips?
This ruined all sorts of single-threaded architectures that relied on these performance increases over time and resulted in programmers being less able to realise performance gains in released chips without having to rework their code to accommodate the change.
That's a "virtue out of necessity" kind of situation -- why would they care for that now that it wont advance as much anymore?
Back in the day though, all the "semiconductor manufacturing" industry touted and advertised strongly was its annual clock frequency advances (similar to how the camera industry switched from touting megapixel advances to low-light performance, video codecs, and other features).
That's Dennard scaling, which indeed failed in the early 2000s. But Moore's law has been quite relevant in other ways, even in the post-Dennard scaling world - even if you discount multiple cores, I don't think "nobody cares about" increased memory/cache sizes, cheap GPU-like accelerators, specialized blocks for e.g. media decompression etc. etc.! If that sort of progress starts slowing down dramatically (and arguably it has in the past few years, especially if you take costs/economic feasibility into account and don't just look at the leading nodes) it's an even bigger change for the industry.
>Hey, just because layman don't care does not mean that there are not people who actually care and use terminology correctly and discuss it.
What matters is the way the grandparent meant it, since that's what we are discussing here, not the etymology or lexicographical definition of the term. That said...
>You should never be proud about dumbing down concepts in Hacker News. This is not consumer forum.
Thanks for the sanctimonious sermon.
If one actually did their research they would find out that it's not just some "layman doesn't care" thing, but that industry insiders, pundits, and others, have used Moore's law in several different ways over the decades, regardless of the original observation by Moore.
Also note that the law was since forever linked with performance increases. Wikipedia: "Moore did not predict a doubling "every 18 months". Rather, David House, an Intel colleague, had factored in the increasing performance of transistors to conclude that integrated circuits would double in performance every 18 months.
Note also despite the name it's neither a physical "law", nor a scientific or technical term or high concept to be closely guarded from laymen. It's just a casual ad-hoc statistical observation.
Not what TFA said, what "Quarrelsome" said that started this subthread:
"I thought Moore's Law effectively ended over a decade ago when we switched to multi-core chips? This ruined all sorts of single-threaded architectures that relied on these performance increases over time and resulted in programmers being less able to realise performance gains in released chips without having to rework their code to accommodate the change."
To which someone pedantically replied that Moore's Law is not about performance increases, but transistor count -- to which I replied that that's just one way it has been used (and not the most popular one either, or the more pertinent to the grandparent's question).
I have a photocopy of a memo from Gordon Moore as part of my computer systems lectures notes that endorses a broader meaning of Moore's law. And more to the point from the descriptivist standpoint, the phrase "Moore's Law" was first used in '75 shortly after people started thinking about Dennard scaling and I don't think there was every a time that the phrase was consistently used only to talk about transistor count since that, in practice for the first 30 years, always went together with transistor density, clock speed, and power efficiency.
Moore observed (what is now called) Dennard scaling in his original paper: that smaller processes allow smaller power consumption (and therefore higher clock speeds).
Depends — “Moore’s Law” is used to describe many different aspects of chip improvement, from “transistors per chip at the current minimum cost per transistor” to clock cycles to absolute GFLOP/s per dollar to computational energy efficiency.
Indeed, what they mean by "Moore's Law" is just the continuing ability to improve processor performance.
They're just saying they haven't hit a brick wall, but the fact is they have hit soft ground that's slowing them down a lot. The actual rate of improvement has been stretching out to a doubling every 2.5 to 3 years and that's likely to get longer still on a long term basis. It is possible some of these new techniques might give us some nice generational boosts from time to time though.
Considering 14nm is nearly 1 year late, and 10nm is nearly 3-4 years late depending on which schedule you looked at. Especially when you consider the absolute majority of shipment from Intel in 2020 will still not be on 10nm. Majority of the Server and Desktop Roadmap, as well as H Series Laptop are still on 14++++.
Did they just restart their Moore's Law counter in 2020?
And for anyone reading and watching the tech space for the past decades, you might notice the sudden increase in Intel's article, information leaks as well as PR pieces in the past 12 months. Just when AMD are doing well. I am not entirely sure this is coincidence.
Rivals might not be the right word, within the 5W profile, Apple, using TSMC's leading edge node has been making chips that exceed Intel x86 in many benchmarks at a significant cost saving.
I don't know if it sounds bright. It sounds like there are developments to be expected in the short term (the way the article presents it sounds like it's just supposed to be good PR for Intel), but it's not clear these tricks scale like process used to scale.
Can you stack three, four, five, ten transistor layers on a chip? And how do you power & cool that stack?
Cooling may be an issue for stacking tons of cpu cores, true. But further increases in cache, RAM, SSD sizes (which may not affect power/cooling as much) will also have benefits.
My understanding is that caches are about as power hungry as the rest of the core, and consume lot of space on the chip, which is why you find large high-TDP desktop & server chips with huge L3s and then mobile chips with rather tiny caches. And huge caches are getting problematic in terms of latency too; if you're doing high performance stuff, you might just have to treat L3 like you used to treat RAM as far as cost of access goes. AIUI we just don't have much to gain on that front unless we can keep actually shrinking the process.
RAM and SSD sizes would mainly benefit from becoming cheaper (although even that's not a given if we're stacking layers and effectively multiplying the area -- with possible defects -- by number of layers) but that unfortunately doesn't translate to performance. Both do also run into thermal limits if you try to push performance. Alternatively you can add lanes for more bandwidth, but that doesn't help with latency, and you need more silicon on the CPU to actually handle it.
Right now I do not have any performance problem that I could solve by throwing more RAM or SSD at it (but I could just walk into a store and buy either and have enough for years to come; I can't walk into a store and buy a CPU that's fast enough to not be a bottleneck for years to come). Where there are bottlenecks, they are due to CPU execution speed or I/O bandwidth & latency.
When i look at how mechanically delicate current cpu sockets are, i see no real reason why some liquid cooling connectors
couldn't be integrated. Which then would connect to some capillary lattice embedded in the substrate. Sure would be more costly(in the beginning), but imagine the possible benefits?
As a rule, Betteridge's Law applies to controversial statements where the author wants to provoke people but doesn't have enough evidence to say it affirmatively. I don't think that applies here.
>Betteridge's Law applies to controversial statements where the author wants to provoke people but doesn't have enough evidence to say it affirmatively
That applies perfectly here though, no?
Moore's law is dead, the statement is controversial, without enough evidence to prove their point and this is mostly a PR piece, as others have pointed out.
This seems like exactly the kind of controversial statement in which the author wants to provoke people by challenging the notion that moore's law is no longer valid.
Moore's law (1965) was that number of transistors per single chip doubles every year. The revised law (1975) is that it doubles every two years.
You can achieve that by increasing chip area, increasing density. In the near future also stacking transistor on the same chip over each other.
Packaging multiple chips into one is not part of the Moore's law. AMD gets better yield by quitting the race and it seems to be real end of Moore's law. Density increase is still going on but it can't keep Moore's law going.
Just putting more and more chips in package would be stupid measure for Moore's law.
Arbitrary sized circuit board can be packaged in epoxy.
The point of chiplets is to make effective use of the leading nodes, obviating yield issues for these while resorting to cheaper, coarser nodes for non-critical parts of the whole package. They're not saying that having a multi-chip package by itself is enough, but this is what's enabling them to leverage other improvements. It's not even restricted to Intel, AMD is doing the same thing.
> Moore's law (1965) was that number of transistors per single chip doubles every year.
Not exactly. It was on the cheapest chip cut.
You can't fix it by trading area for yield, nor by creating more expensive denser processes. Chip manufacturers are currently doing both of those, what is good, obviously, but won't bring the kind of evolution we used to get.
I assume that the "cheapest chip cut" is also improving over time, if only in the sense of cost per unit of transistor area. That's perhaps the facet of Moore's law that's going to be comparatively easiest to keep going.
Moore’s law is dying. The transistor and clock speed improvements we saw with regular cycles (33 - 66 - 133 - 266 - 433 - 1Ghz!) hasn’t happened in a while, and throughput improvements have happened by architectural cleverness (some of which you might say contributed to spectre/ meltdown et al).
A really really interesting corollary goes like this:
- Moore’s law delivers a doubling of compute capacity every 2 yrs
- capacity yields efficiency improvements in compute capacity per human being
- more efficiency per human yields net productivity gains per human
- net productivity gains drive economic growth
- therefore Moore’s law or something like it is a critical driver of economic growth
This is a bit scary of an implication - and why as an industry we’re highly incented to come up with something to keep feeding the masses. If we don’t - a key driver of worldwide economic growth will die.
interesting that the article says they’ll keep delivering on the promise - We have to - but is repaint the picture and say that we’re obligated to do so via more exotic software and hardware architectures, so expect to see more purpose built compute in all arenas
Moore's law is not about doubling clock speed, it is about doubling the transistor density.
This has continued after clock speed improvements have stopped.
We're going to have to relearn how to wring every bit of performance out of the chips we have. I'm bearish on the current trend of bloatier and bloatier runtime environments; stuffing an entire isolated instance of Chrome in a container to run a Todo-list app shouldn't have ever made sense, but it especially does not in an environment where compute and memory is once again precious.
I hope that betting on Rust over JavaScript would be the winning play over the next decade, but I'll probably be wrong.
Certainly many kinds of apps (e.g. web- and GUI- and disk- or DB-based) won't care if Dennard scaling has essentially stopped on a uniprocessor. But in other domains uniprocess speed is still their life's blood, like graphics, AR/VR, data-parallel number crunching, secure computing, and increasingly I suspect, non-deepNN forms of AI (like search).
As clock rates fall further behind the venerable curve of CPU clock rate doubling, I too wonder how long the 10x to 30x slowdown obliged by interpreted languages can last. It's lovely to write code a bit faster using REPL. But if that code's slow runtime or likelier troubles with portability or error recovery diminish its utility, then that bargain is Faustian.
I see some of that already after talking to a few startups who are in need low level coding abilities. These days there seems to be a deficit of folks with proficiency in fare like the GCC toolchain, that is, at the levels of assembler, binary objects and libraries, linkers, and device-drivers.
I wonder, given the proclivity of CS academia for the past 20 years to ground their instructional SW concepts mostly at a higher levels than binary (like reliance on libraries, OOPL objects, and pseudocode), if this canny valley is likely to grow into a real obstruction -- esp. now that we likely can afford it least.
Your best bet for future support for efficient code is a language that emphasizes support for coding powerful, efficient libraries that can be used anywhere, for any purpose, without compromise. Such libraries earn the extensive testing and optimization that widespread usage both demands and justifies. High-level, efficient libraries enable coding at a high level, avoiding low-level pitfalls.
C++ is still all alone in pursuing this course, today. Rust might start off in that direction after it matures more.
A very interesting point I heard in a seminar* about Moore's law is that is was before all a commercial roadmap: at that point, Gordon Moore was probably more business-oriented than research-oriented, and such predictable improvements are above all useful to plan investments, so it seems that the semiconductor industry tried very hard to remain on that roadmap all these years (and obviously failed recently).
* by Jean-Pierre Raskin (UC Louvain) about what would the semiconductor industry look like tomorrow, and how to make it more responsible.
>At its simplest level, Moore’s Law refers to a doubling of transistors on a chip with each process generation.
That is an interesting way of moving the goalposts. The actual observation is that the transistors will double "about every two years."
With Intel moving from a tick-tock to a tick-tick-tick-tick-tick-tick-tick-maybetock I can see why they want to redefine Moore's Law to reference their new reality.
Yeah, wow, that's egregious. And now the vague "system scaling" also keeps this Intel-definition of "Moore's Law" alive and well. Clearly Intel are setting up their own hoops to jump through.
> System scaling improvements are the gains that help us incorporate new types of heterogeneous processors via advances in chiplets, packaging, and high-bandwidth chip-to-chip interconnect technologies
There has never been any other product in modern history, or possibly in the entire human history that has advanced as fast as transistor fabrication. Moore's law has been doubling transistor every 2 years for the past 50 years from 1965 to 2015. It wasn't until then we hit a hiccup.
Not sleeping leads to death in remarkably short order. (Nobody knows why, but the fact is well established.) Dying cannot be expected to yield faster computers.
In China they build entire skyscrapers within days, by working on it 24/7. Apparently it's possible to switch out workers. We don't build buildings that fast over here, because reasons.
I'm wondering if our computers ain't five times faster, because of reasons, too.
51 comments
[ 3.6 ms ] story [ 109 ms ] threadThis ruined all sorts of single-threaded architectures that relied on these performance increases over time and resulted in programmers being less able to realise performance gains in released chips without having to rework their code to accommodate the change.
What people cared about, and referred to as "Moore's law" was the cheap automatic doubling of performance (mostly due to clock frequency increases).
That's long gone with multicore...
Back in the day though, all the "semiconductor manufacturing" industry touted and advertised strongly was its annual clock frequency advances (similar to how the camera industry switched from touting megapixel advances to low-light performance, video codecs, and other features).
You should never be proud about dumbing down concepts in Hacker News. This is not consumer forum.
What matters is the way the grandparent meant it, since that's what we are discussing here, not the etymology or lexicographical definition of the term. That said...
>You should never be proud about dumbing down concepts in Hacker News. This is not consumer forum.
Thanks for the sanctimonious sermon.
If one actually did their research they would find out that it's not just some "layman doesn't care" thing, but that industry insiders, pundits, and others, have used Moore's law in several different ways over the decades, regardless of the original observation by Moore.
Also note that the law was since forever linked with performance increases. Wikipedia: "Moore did not predict a doubling "every 18 months". Rather, David House, an Intel colleague, had factored in the increasing performance of transistors to conclude that integrated circuits would double in performance every 18 months.
Note also despite the name it's neither a physical "law", nor a scientific or technical term or high concept to be closely guarded from laymen. It's just a casual ad-hoc statistical observation.
It's all about complexity per chip. Complexity as number of components per chip.
"I thought Moore's Law effectively ended over a decade ago when we switched to multi-core chips? This ruined all sorts of single-threaded architectures that relied on these performance increases over time and resulted in programmers being less able to realise performance gains in released chips without having to rework their code to accommodate the change."
To which someone pedantically replied that Moore's Law is not about performance increases, but transistor count -- to which I replied that that's just one way it has been used (and not the most popular one either, or the more pertinent to the grandparent's question).
They're just saying they haven't hit a brick wall, but the fact is they have hit soft ground that's slowing them down a lot. The actual rate of improvement has been stretching out to a doubling every 2.5 to 3 years and that's likely to get longer still on a long term basis. It is possible some of these new techniques might give us some nice generational boosts from time to time though.
Considering 14nm is nearly 1 year late, and 10nm is nearly 3-4 years late depending on which schedule you looked at. Especially when you consider the absolute majority of shipment from Intel in 2020 will still not be on 10nm. Majority of the Server and Desktop Roadmap, as well as H Series Laptop are still on 14++++.
Did they just restart their Moore's Law counter in 2020?
And for anyone reading and watching the tech space for the past decades, you might notice the sudden increase in Intel's article, information leaks as well as PR pieces in the past 12 months. Just when AMD are doing well. I am not entirely sure this is coincidence.
Can you stack three, four, five, ten transistor layers on a chip? And how do you power & cool that stack?
RAM and SSD sizes would mainly benefit from becoming cheaper (although even that's not a given if we're stacking layers and effectively multiplying the area -- with possible defects -- by number of layers) but that unfortunately doesn't translate to performance. Both do also run into thermal limits if you try to push performance. Alternatively you can add lanes for more bandwidth, but that doesn't help with latency, and you need more silicon on the CPU to actually handle it.
Right now I do not have any performance problem that I could solve by throwing more RAM or SSD at it (but I could just walk into a store and buy either and have enough for years to come; I can't walk into a store and buy a CPU that's fast enough to not be a bottleneck for years to come). Where there are bottlenecks, they are due to CPU execution speed or I/O bandwidth & latency.
So I guess the answer is no.
That applies perfectly here though, no?
Moore's law is dead, the statement is controversial, without enough evidence to prove their point and this is mostly a PR piece, as others have pointed out.
Yeah, I can see why Intel would want you to believe that.
You can achieve that by increasing chip area, increasing density. In the near future also stacking transistor on the same chip over each other.
Packaging multiple chips into one is not part of the Moore's law. AMD gets better yield by quitting the race and it seems to be real end of Moore's law. Density increase is still going on but it can't keep Moore's law going.
Just putting more and more chips in package would be stupid measure for Moore's law. Arbitrary sized circuit board can be packaged in epoxy.
Not exactly. It was on the cheapest chip cut.
You can't fix it by trading area for yield, nor by creating more expensive denser processes. Chip manufacturers are currently doing both of those, what is good, obviously, but won't bring the kind of evolution we used to get.
But it has always been a minor gain. Boubling it every few years will stop being useful very quickly.
A really really interesting corollary goes like this: - Moore’s law delivers a doubling of compute capacity every 2 yrs
- capacity yields efficiency improvements in compute capacity per human being
- more efficiency per human yields net productivity gains per human
- net productivity gains drive economic growth
- therefore Moore’s law or something like it is a critical driver of economic growth
This is a bit scary of an implication - and why as an industry we’re highly incented to come up with something to keep feeding the masses. If we don’t - a key driver of worldwide economic growth will die.
interesting that the article says they’ll keep delivering on the promise - We have to - but is repaint the picture and say that we’re obligated to do so via more exotic software and hardware architectures, so expect to see more purpose built compute in all arenas
Even today it is following Moore's law, although not at Intel: https://medium.com/predict/moores-law-is-alive-and-well-eaa4...
I hope that betting on Rust over JavaScript would be the winning play over the next decade, but I'll probably be wrong.
As clock rates fall further behind the venerable curve of CPU clock rate doubling, I too wonder how long the 10x to 30x slowdown obliged by interpreted languages can last. It's lovely to write code a bit faster using REPL. But if that code's slow runtime or likelier troubles with portability or error recovery diminish its utility, then that bargain is Faustian.
I see some of that already after talking to a few startups who are in need low level coding abilities. These days there seems to be a deficit of folks with proficiency in fare like the GCC toolchain, that is, at the levels of assembler, binary objects and libraries, linkers, and device-drivers.
I wonder, given the proclivity of CS academia for the past 20 years to ground their instructional SW concepts mostly at a higher levels than binary (like reliance on libraries, OOPL objects, and pseudocode), if this canny valley is likely to grow into a real obstruction -- esp. now that we likely can afford it least.
C++ is still all alone in pursuing this course, today. Rust might start off in that direction after it matures more.
* by Jean-Pierre Raskin (UC Louvain) about what would the semiconductor industry look like tomorrow, and how to make it more responsible.
That is an interesting way of moving the goalposts. The actual observation is that the transistors will double "about every two years."
With Intel moving from a tick-tock to a tick-tick-tick-tick-tick-tick-tick-maybetock I can see why they want to redefine Moore's Law to reference their new reality.
> System scaling improvements are the gains that help us incorporate new types of heterogeneous processors via advances in chiplets, packaging, and high-bandwidth chip-to-chip interconnect technologies
I'm wondering if our computers ain't five times faster, because of reasons, too.