11 ms·

ASML looking to finally get ahead of Moore’s Law

180 points by kristianp 4y ago

zogomoox 4y ago
In development High-NA EUV can do 2nm logic, possibly 1nm with double patterning? There's not a lot more room at the bottom after that, given silicon crystal atom distance of .3nm
- subaken 4y ago
  "2nm" is a marketing term. There are no actual parts that actually measure 2nm.
  - OhNoNotAgain_99 4y ago
    It's wire width the finest wires doped positive or negative do the job. So a transistor is some 3d layered construct based on those wires
- thechao 4y ago
  2nm transistor is still on the order of 10s of millions of atoms.
  - px43 4y ago
    Silicon atoms are 0.2nm across. 2nm is 10 atoms wide.
    - adrian_b 4y ago
      I do not know from where you have got that number, but it is meaningless. The silicon atoms do not behave like spheres, only the atoms from many metallic or ionic substances have a behavior close to that of spheres. So it makes no sense to speak about a length "across" a silicon atom. In the case of silicon, a meaningful number is that there are around 50 silicon atoms per cubic nanometer of silicon crystal. The gate of a transistor (which is the active part of the transistor, and which has a much smaller volume than the complete transistor) in the latest 5-nm technologies is contained in a fin that might have a width around 5 nm, a length around 20 nm and a height around 50 nm, for a volume around 5000 cubic namometers (it is a complete coincidence that the width of a fin might be around 5 nm to 6 nm for a process named as "5 nm"; there is no relationship between the name of the process and the width of the fin; fins did not even exist for processes with names greater than "22 nm"). Such a gate of the smallest transistor might include around 250 thousand Si atoms. The volume of the complete transistor would be at least 10 to 20 times greater. However the size of a transistor is not limited by the number of silicon atoms in the gate, but by the number of impurity atoms that control the conductivity of the silicon, and those are much less than the silicon atoms (I have not seen any number for the latest technologies, but they could be e.g. 10 thousand times less than the silicon atoms, so there might be less than 100 impurity atoms in the gate).
      - benj111 4y ago
        >I do not know from where you have got that number, but it is meaningless Could you expand a bit? If atoms were randomly arranged, or in a non solid state I could kind of understand. But once you get to any kind of fixed structure you can infer 2d spacing from 3d spacing. Or is your intent to say that the quantity of silicon in a gate is 3d, especially with fins, so a 2d view doesn't give a complete picture?
        adrian_b 4y ago
        In silicon crystals and in most other semiconductors, the atoms are kept in their positions by covalent bonds that have certain directions in space, e.g. towards the 4 vertices of a regular tetrahedron in the case of a silicon crystal. Because of that, the atoms are not packed together like some spheres, i.e. like the atoms in metallic aluminum or in table salt, where you may speak about the diameter of the atomic spheres. They are distributed on a lattice that has empty spaces between atoms and their bonds (i.e. the places where electrons belonging to the atoms are located with high probability). The distance between the silicon atoms in a silicon crystal varies depending on the direction, so there is no single value that could be considered the diameter of a silicon atom. The periodic cell of a silicon crystal has the same structure as that of cubic diamond and it has the form of a cube with 8 atoms inside it (an atom in a cube corner counts as 1/8 inside, an atom on a face counts as 1/2 inside). A visualization from Wikipedia: https://en.wikipedia.org/wiki/Diamond_cubic#/media/File:Diamond_cubic_animation.gif https://en.wikipedia.org/wiki/Diamond_cubic#/media/File:Diam... While this visualization uses balls and sticks, to show the positions of the centers of the atoms, that has nothing to do with the form of the real atoms. At most you could consider that a silicon atom has the form of the corresponding Voronoi polyhedron, in which case you would have to give several numbers, to describe its size, and not a single "across" value: https://en.wikipedia.org/wiki/Triakis_truncated_tetrahedron https://en.wikipedia.org/wiki/Triakis_truncated_tetrahedron The number that I have provided, i.e. 50 Si atoms per cubic nanometer of Si crystal, can be computed by dividing 8 atoms to the volume of the cubic cell of the Si lattice. Given a volume of Si crystal, you can compute the number of Si atoms. Like I have said, I do not know what means that 0.2 nm value, as the distance between 2 neighbor Si atoms can be larger than 0.5 nm, depending on the direction. In any case you cannot use it to compute anything about the number of atoms in a silicon device. EDIT: I believe that you might have got your 0.2 nm from truncating the distance between 2 silicon atomic planes in the so-called "111" direction (the direction of the cube diagonal), which is the minimum distance between atomic planes in silicon. That distance is 0.543 nm * sqrt(3) / 4 = 0.235 nm. Because this distance is correct only for the "111" crystalographic direction, it cannot be used to compute the number of atoms in some piece of silicon, and it certainly cannot be called as the diameter of a silicon atom ("across an atom" without specifying the direction).
    - thechao 4y ago
      adrian_b's explanation is pretty good explanation of what I'm saying. To expand on that: you cannot pack transistors side-by-side without moving to global custom poly. The result is that there's a large amount of "technically not space for the transistor, but space required for the transistor to work" stuff around the transistor. The result is that a 2nm process transistors is on the order of 10s of millions of atoms. Source: I work with the latest nodes. Here's a Fermi estimate: the M1 is ~16 billion transistors; assume it is 1 cm^2. That gives (10^7)^2 nm^2 / 10^10 transistors, which is ~10^4 nm^2/transistor. Assume the transistor is ~50nm high, which gives ~5x10^5 nm^3/transistor. There are ~100 SI atoms/nm^3, which gives a volume of ~5x10^7 SI/transistor.
drexlspivey 4y ago
ASML recently announced a massive stock buyback in the order of $12B (55% of their revenue) which left me scratching my head. They reportedly can’t meet demand and the current backlog for new machines is at 2+ years. Surely a company in a sector with such massive capital expenditures has better use of this money? For example scaling production or R&D?
- phkahler 4y ago
  My guess is they are spending R&D money on better/simpler/cheaper EUV light sources, so expanding production capacity of existing models isn't a great idea. Specifically my guess (and I'm no expert) is a plasma wakefield accelerator feeding a free electron laser to get tunable wavelength EUV light far cheaper and simpler than what they've got now. Tons of work to do to get there, but that's always been the case in thier business. Also, if they don't do it someone else might and then they'd be dead. There are other R&D areas to be handled too. One might ask why not spend more on the R&D, but they are probably at some kind of practical limit there too.
  - usehand 4y ago
    > Specifically my guess (and I'm no expert) is a plasma wakefield accelerator feeding a free electron laser to get tunable wavelength EUV light far cheaper and simpler than what they've got now Don't you think they'll also be using a retro encabulator, with the modial interaction of magneto-reluctance and capacitive directance being used to generate the EUV light?
    - throwaway773748 4y ago
      I honestly can not tell if this is a serious question or a joke throwing around physics techno babble... Maybe this is how business people feel when we try to explain why some data somewhere is out of sync because of system problems.
      - usehand 4y ago
        https://www.youtube.com/watch?v=RXJKdh1KZ0w https://www.youtube.com/watch?v=RXJKdh1KZ0w (:
1letterunixname 4y ago
Expecting hockey stick growth indefinitely is patently ridiculous. Surely there will be an inevitable insurmountable plateau reducing cost and performance gains to smaller improvements. At some point, different technologies will have to be seriously considered such as photonics.
- insanitybit 4y ago
  Aren't there still massive gains on the table? I'm totally naive on this but I recall hearing about other materials, such as graphite iirc, that could achieve 100x CPU hz.
  - OhNoNotAgain_99 4y ago
  - dataking 4y ago
    According to Jim Keller, yes, there is. If my memory serves, he said so in a recent Lex Friedman interview. Edit: here is the clip https://youtu.be/c01BlUDIlK4 https://youtu.be/c01BlUDIlK4
    - insanitybit 4y ago
      Thanks, this is pretty much what I was getting at. While "Moore's Law" might strictly mean transistor density, "performance" is about many things.
  - FeepingCreature 4y ago
    We're already limited by lightspeed propagation, so it's hard to see how this could be the case.
    - insanitybit 4y ago
      I assume what you mean is that the CPU may get faster but would ultimately stall waiting for data, which can only move so fast. Presumably to continue feeding data to the CPU we'd want to be much more aggressive about caching and prefetching. Another option would be to not go 100x faster but to go 10x faster and then take the improved cooling + energy usage as a win.
XCSme 4y ago
I was checking the ASML technology page[0] as I haven't heard of them before, it was funny seeing that they are a bleeding-edge tech company and they have an image showing jQuery code next to a paragraph about their software: https://i.snipboard.io/pUT4ot.jpg https://i.snipboard.io/pUT4ot.jpg [0]: https://www.asml.com/en/technology https://www.asml.com/en/technology
- troymc 4y ago
  I'm guessing that the page designer just went to a stock photo site, did a search for "code" or something similar, found that image, licensed it, and inserted it. Job done! I asked TinEye [1] where I can get that stock photo. It turns out you can get it at most of the big stock photo sites. For example, here it is on Shutterstock: https://www.shutterstock.com/image-photo/software-source-code-programming-on-computer-634574354 https://www.shutterstock.com/image-photo/software-source-cod... [1] https://tineye.com/ https://tineye.com/
- jimmaswell 4y ago
  There's nothing wrong with using jQuery, especially if you target old browsers or use some of its features that are still more convenient than the ES6 equivalent.
  - XCSme 4y ago
    Agree, but I don't think that jQuery really represents the core of their software, I find it funny how big the discrepancy between the text and the image is.
- yucky 4y ago
  Even calling them a bleeding edge tech company is selling them short. They are the only company in the world that can do what they do. Their website could have pictures of clowns in cars on it, and it wouldn't matter.
- baybal2 4y ago
- 4y ago
tromp 4y ago
> high-NA EUV TIL that NA stands for numerical aperture, which is a measure of how much light the optical system can collect and focus.
- zh3 4y ago
  It's actually a measure of the angle over which light can be collected, and it's defined as the sine of the half-angle. For example, if you want to launch light into a fibre-optic cable with NA 0.5, the light needs to be angled 30 degrees or less with respect to the cable (since sin(30) = 0./5); any light with a greater angle won't undergo reflection and so simply shoot out the side of the fibre (conversely, light at smaller angles will be reflected and so go bouncing down the cable until it comes out at the far end - again with a 30 degree half-angle).
bullen 4y ago
While you are debating if eternal growth exists we that actually write software can tell you that the bottle-neck has been RAM since 386 that introduced caches: DDR3 was the peak of RAM latency, 65nm SSDs where the peak of flash memory longevity both in ~2010! You cannot scale anything unless it's embarrasingly parallelizable in which case you never had a problem in the first place. Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W). Everything that is below one order of magnitude improvement is irrelevant = ASML is irrelevant unless devices break and you buy things you cannot repair. That combined with eternally raising electricity cost, not as a function of marginal prices on a market (like today), but as a permanent depletion of resources leads to only 1 conclusion: Don't wait for the next generation hardware/firmware locked, rentseeking line of hardware: 14nm X86 and 28nm ARM IS good enough for a lifetime if you control it and make sure it lasts. There are only 2 programming languages that can survive this long term: C on the client (with a little ++ for convenience) and vanilla JavaSE on the server (with as little deps. as possible). Bonus prediction: JavaME will be revived in some form for open microcontroller mobile use.
- FridgeSeal 4y ago
  > There are only 2 programming languages that can survive this long term: C on the client (with a little ++ for convenience) and vanilla JavaSE That’s kind of a big claim, why only these 2?
  - IshKebab 4y ago
    I mean it's clearly nonsense. Why even ask?
- awestroke 4y ago
  This reads like a TimeCube rant
- account42 4y ago
  > Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W). Those numbers are very misleading - perf per watt is not a constant but something that very much depends on how far the clock speed is pushed. Anything tuned for perf is going to have significantly worse perf per watt than the architecture is capable of. The same goes for the CPU design - maximizing perf and maximizing perf per watt is going to give you different architectures on the same node.
scotty79 4y ago
Is there an equivalent of Moore's law for energy efficiency of computation?
- bjornsing 4y ago
  Not that I know of. But interestingly there is a theoretical limit. One implication of that is of course that there can’t be an open ended equivalent of Moore’s law; any such law must come to end.
- fisf 4y ago
  https://en.m.wikipedia.org/wiki/Landauer%27s_principle https://en.m.wikipedia.org/wiki/Landauer%27s_principle
  - scotty79 4y ago
    I meant something like this: https://www.technologyreview.com/2011/09/12/191382/a-new-and-improved-moores-law/ https://www.technologyreview.com/2011/09/12/191382/a-new-and...
simne 4y ago
This looks crazy! Remembers me madness on plastic windows in Ukraine, where lot of people buy machine-tools, with dreams, to sell plastic windows forever. Some even build full size factories. But unfortunately all these ended, because hundreds of small businesses in ~10 years, filled all current needs of people, and plastic windows last about 20 years, so once market shrink by about 95%, because now all these tools will do 1/20 of market per year. PS sure, war will change things, but even most pessimistic opinions, about 30% of homes need big repair (really much less, but now impossible to know exactly), and with existing capacity, they will be filled in cup of years.
simne 4y ago
BTW, if things really so good, I hear, CANON (yes the same who make photo cameras) works on steppers, and one other Japanese company, forgot name, subsidiary of one known big group. This is not easy business, but there is space for at least one more player.

180 points by kristianp 4y ago

In development High-NA EUV can do 2nm logic, possibly 1nm with double patterning? There's not a lot more room at the bottom after that, given silicon crystal atom distance of .3nm

subaken 4y ago
"2nm" is a marketing term. There are no actual parts that actually measure 2nm.
- OhNoNotAgain_99 4y ago
  It's wire width the finest wires doped positive or negative do the job. So a transistor is some 3d layered construct based on those wires
thechao 4y ago
2nm transistor is still on the order of 10s of millions of atoms.
- px43 4y ago
  Silicon atoms are 0.2nm across. 2nm is 10 atoms wide.
  - adrian_b 4y ago
    I do not know from where you have got that number, but it is meaningless. The silicon atoms do not behave like spheres, only the atoms from many metallic or ionic substances have a behavior close to that of spheres. So it makes no sense to speak about a length "across" a silicon atom. In the case of silicon, a meaningful number is that there are around 50 silicon atoms per cubic nanometer of silicon crystal. The gate of a transistor (which is the active part of the transistor, and which has a much smaller volume than the complete transistor) in the latest 5-nm technologies is contained in a fin that might have a width around 5 nm, a length around 20 nm and a height around 50 nm, for a volume around 5000 cubic namometers (it is a complete coincidence that the width of a fin might be around 5 nm to 6 nm for a process named as "5 nm"; there is no relationship between the name of the process and the width of the fin; fins did not even exist for processes with names greater than "22 nm"). Such a gate of the smallest transistor might include around 250 thousand Si atoms. The volume of the complete transistor would be at least 10 to 20 times greater. However the size of a transistor is not limited by the number of silicon atoms in the gate, but by the number of impurity atoms that control the conductivity of the silicon, and those are much less than the silicon atoms (I have not seen any number for the latest technologies, but they could be e.g. 10 thousand times less than the silicon atoms, so there might be less than 100 impurity atoms in the gate).
    - benj111 4y ago
      >I do not know from where you have got that number, but it is meaningless Could you expand a bit? If atoms were randomly arranged, or in a non solid state I could kind of understand. But once you get to any kind of fixed structure you can infer 2d spacing from 3d spacing. Or is your intent to say that the quantity of silicon in a gate is 3d, especially with fins, so a 2d view doesn't give a complete picture?
      - adrian_b 4y ago
        In silicon crystals and in most other semiconductors, the atoms are kept in their positions by covalent bonds that have certain directions in space, e.g. towards the 4 vertices of a regular tetrahedron in the case of a silicon crystal. Because of that, the atoms are not packed together like some spheres, i.e. like the atoms in metallic aluminum or in table salt, where you may speak about the diameter of the atomic spheres. They are distributed on a lattice that has empty spaces between atoms and their bonds (i.e. the places where electrons belonging to the atoms are located with high probability). The distance between the silicon atoms in a silicon crystal varies depending on the direction, so there is no single value that could be considered the diameter of a silicon atom. The periodic cell of a silicon crystal has the same structure as that of cubic diamond and it has the form of a cube with 8 atoms inside it (an atom in a cube corner counts as 1/8 inside, an atom on a face counts as 1/2 inside). A visualization from Wikipedia: https://en.wikipedia.org/wiki/Diamond_cubic#/media/File:Diamond_cubic_animation.gif https://en.wikipedia.org/wiki/Diamond_cubic#/media/File:Diam... While this visualization uses balls and sticks, to show the positions of the centers of the atoms, that has nothing to do with the form of the real atoms. At most you could consider that a silicon atom has the form of the corresponding Voronoi polyhedron, in which case you would have to give several numbers, to describe its size, and not a single "across" value: https://en.wikipedia.org/wiki/Triakis_truncated_tetrahedron https://en.wikipedia.org/wiki/Triakis_truncated_tetrahedron The number that I have provided, i.e. 50 Si atoms per cubic nanometer of Si crystal, can be computed by dividing 8 atoms to the volume of the cubic cell of the Si lattice. Given a volume of Si crystal, you can compute the number of Si atoms. Like I have said, I do not know what means that 0.2 nm value, as the distance between 2 neighbor Si atoms can be larger than 0.5 nm, depending on the direction. In any case you cannot use it to compute anything about the number of atoms in a silicon device. EDIT: I believe that you might have got your 0.2 nm from truncating the distance between 2 silicon atomic planes in the so-called "111" direction (the direction of the cube diagonal), which is the minimum distance between atomic planes in silicon. That distance is 0.543 nm * sqrt(3) / 4 = 0.235 nm. Because this distance is correct only for the "111" crystalographic direction, it cannot be used to compute the number of atoms in some piece of silicon, and it certainly cannot be called as the diameter of a silicon atom ("across an atom" without specifying the direction).
  - thechao 4y ago
    adrian_b's explanation is pretty good explanation of what I'm saying. To expand on that: you cannot pack transistors side-by-side without moving to global custom poly. The result is that there's a large amount of "technically not space for the transistor, but space required for the transistor to work" stuff around the transistor. The result is that a 2nm process transistors is on the order of 10s of millions of atoms. Source: I work with the latest nodes. Here's a Fermi estimate: the M1 is ~16 billion transistors; assume it is 1 cm^2. That gives (10^7)^2 nm^2 / 10^10 transistors, which is ~10^4 nm^2/transistor. Assume the transistor is ~50nm high, which gives ~5x10^5 nm^3/transistor. There are ~100 SI atoms/nm^3, which gives a volume of ~5x10^7 SI/transistor.

drexlspivey 4y ago

ASML recently announced a massive stock buyback in the order of $12B (55% of their revenue) which left me scratching my head. They reportedly can’t meet demand and the current backlog for new machines is at 2+ years. Surely a company in a sector with such massive capital expenditures has better use of this money? For example scaling production or R&D?

phkahler 4y ago
My guess is they are spending R&D money on better/simpler/cheaper EUV light sources, so expanding production capacity of existing models isn't a great idea. Specifically my guess (and I'm no expert) is a plasma wakefield accelerator feeding a free electron laser to get tunable wavelength EUV light far cheaper and simpler than what they've got now. Tons of work to do to get there, but that's always been the case in thier business. Also, if they don't do it someone else might and then they'd be dead. There are other R&D areas to be handled too. One might ask why not spend more on the R&D, but they are probably at some kind of practical limit there too.
- usehand 4y ago
  > Specifically my guess (and I'm no expert) is a plasma wakefield accelerator feeding a free electron laser to get tunable wavelength EUV light far cheaper and simpler than what they've got now Don't you think they'll also be using a retro encabulator, with the modial interaction of magneto-reluctance and capacitive directance being used to generate the EUV light?
  - throwaway773748 4y ago
    I honestly can not tell if this is a serious question or a joke throwing around physics techno babble... Maybe this is how business people feel when we try to explain why some data somewhere is out of sync because of system problems.
    - usehand 4y ago
      https://www.youtube.com/watch?v=RXJKdh1KZ0w https://www.youtube.com/watch?v=RXJKdh1KZ0w (:

1letterunixname 4y ago

Expecting hockey stick growth indefinitely is patently ridiculous. Surely there will be an inevitable insurmountable plateau reducing cost and performance gains to smaller improvements. At some point, different technologies will have to be seriously considered such as photonics.

insanitybit 4y ago
Aren't there still massive gains on the table? I'm totally naive on this but I recall hearing about other materials, such as graphite iirc, that could achieve 100x CPU hz.
- OhNoNotAgain_99 4y ago
- dataking 4y ago
  According to Jim Keller, yes, there is. If my memory serves, he said so in a recent Lex Friedman interview. Edit: here is the clip https://youtu.be/c01BlUDIlK4 https://youtu.be/c01BlUDIlK4
  - insanitybit 4y ago
    Thanks, this is pretty much what I was getting at. While "Moore's Law" might strictly mean transistor density, "performance" is about many things.
- FeepingCreature 4y ago
  We're already limited by lightspeed propagation, so it's hard to see how this could be the case.
  - insanitybit 4y ago
    I assume what you mean is that the CPU may get faster but would ultimately stall waiting for data, which can only move so fast. Presumably to continue feeding data to the CPU we'd want to be much more aggressive about caching and prefetching. Another option would be to not go 100x faster but to go 10x faster and then take the improved cooling + energy usage as a win.

XCSme 4y ago

I was checking the ASML technology page[0] as I haven't heard of them before, it was funny seeing that they are a bleeding-edge tech company and they have an image showing jQuery code next to a paragraph about their software: https://i.snipboard.io/pUT4ot.jpg https://i.snipboard.io/pUT4ot.jpg [0]: https://www.asml.com/en/technology https://www.asml.com/en/technology

troymc 4y ago
I'm guessing that the page designer just went to a stock photo site, did a search for "code" or something similar, found that image, licensed it, and inserted it. Job done! I asked TinEye [1] where I can get that stock photo. It turns out you can get it at most of the big stock photo sites. For example, here it is on Shutterstock: https://www.shutterstock.com/image-photo/software-source-code-programming-on-computer-634574354 https://www.shutterstock.com/image-photo/software-source-cod... [1] https://tineye.com/ https://tineye.com/
jimmaswell 4y ago
There's nothing wrong with using jQuery, especially if you target old browsers or use some of its features that are still more convenient than the ES6 equivalent.
- XCSme 4y ago
  Agree, but I don't think that jQuery really represents the core of their software, I find it funny how big the discrepancy between the text and the image is.
yucky 4y ago
Even calling them a bleeding edge tech company is selling them short. They are the only company in the world that can do what they do. Their website could have pictures of clowns in cars on it, and it wouldn't matter.
baybal2 4y ago
4y ago

tromp 4y ago

> high-NA EUV TIL that NA stands for numerical aperture, which is a measure of how much light the optical system can collect and focus.

zh3 4y ago
It's actually a measure of the angle over which light can be collected, and it's defined as the sine of the half-angle. For example, if you want to launch light into a fibre-optic cable with NA 0.5, the light needs to be angled 30 degrees or less with respect to the cable (since sin(30) = 0./5); any light with a greater angle won't undergo reflection and so simply shoot out the side of the fibre (conversely, light at smaller angles will be reflected and so go bouncing down the cable until it comes out at the far end - again with a 30 degree half-angle).

bullen 4y ago

While you are debating if eternal growth exists we that actually write software can tell you that the bottle-neck has been RAM since 386 that introduced caches: DDR3 was the peak of RAM latency, 65nm SSDs where the peak of flash memory longevity both in ~2010! You cannot scale anything unless it's embarrasingly parallelizable in which case you never had a problem in the first place. Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W). Everything that is below one order of magnitude improvement is irrelevant = ASML is irrelevant unless devices break and you buy things you cannot repair. That combined with eternally raising electricity cost, not as a function of marginal prices on a market (like today), but as a permanent depletion of resources leads to only 1 conclusion: Don't wait for the next generation hardware/firmware locked, rentseeking line of hardware: 14nm X86 and 28nm ARM IS good enough for a lifetime if you control it and make sure it lasts. There are only 2 programming languages that can survive this long term: C on the client (with a little ++ for convenience) and vanilla JavaSE on the server (with as little deps. as possible). Bonus prediction: JavaME will be revived in some form for open microcontroller mobile use.

FridgeSeal 4y ago
> There are only 2 programming languages that can survive this long term: C on the client (with a little ++ for convenience) and vanilla JavaSE That’s kind of a big claim, why only these 2?
- IshKebab 4y ago
  I mean it's clearly nonsense. Why even ask?
awestroke 4y ago
This reads like a TimeCube rant
account42 4y ago
> Peak performance is more important than the puny 0.5Gflops/W "5"nm M1 (2.5Gflops/W) has over the 28nm Raspberry 4 (2 Gflops/W). Those numbers are very misleading - perf per watt is not a constant but something that very much depends on how far the clock speed is pushed. Anything tuned for perf is going to have significantly worse perf per watt than the architecture is capable of. The same goes for the CPU design - maximizing perf and maximizing perf per watt is going to give you different architectures on the same node.

scotty79 4y ago

Is there an equivalent of Moore's law for energy efficiency of computation?

bjornsing 4y ago
Not that I know of. But interestingly there is a theoretical limit. One implication of that is of course that there can’t be an open ended equivalent of Moore’s law; any such law must come to end.
fisf 4y ago
https://en.m.wikipedia.org/wiki/Landauer%27s_principle https://en.m.wikipedia.org/wiki/Landauer%27s_principle
- scotty79 4y ago
  I meant something like this: https://www.technologyreview.com/2011/09/12/191382/a-new-and-improved-moores-law/ https://www.technologyreview.com/2011/09/12/191382/a-new-and...

simne 4y ago

This looks crazy! Remembers me madness on plastic windows in Ukraine, where lot of people buy machine-tools, with dreams, to sell plastic windows forever. Some even build full size factories. But unfortunately all these ended, because hundreds of small businesses in ~10 years, filled all current needs of people, and plastic windows last about 20 years, so once market shrink by about 95%, because now all these tools will do 1/20 of market per year. PS sure, war will change things, but even most pessimistic opinions, about 30% of homes need big repair (really much less, but now impossible to know exactly), and with existing capacity, they will be filled in cup of years.

simne 4y ago

BTW, if things really so good, I hear, CANON (yes the same who make photo cameras) works on steppers, and one other Japanese company, forgot name, subsidiary of one known big group. This is not easy business, but there is space for at least one more player.