28 ms·

No More Nanometers – It’s Time for New Node Naming (2020)

121 points by bcaa7f3a8bbc 5y ago

aappleby 5y ago
no mention of https://en.wikichip.org/wiki/mtr-mm%C2%B2 https://en.wikichip.org/wiki/mtr-mm%C2%B2 ?
- Waterluvian 5y ago
  I know very little about this, but shouldn't the ideal unit consider volume? They're going to eventually do clever 3d stuff I'm guessing.
  - wtallis 5y ago
    CPUs are starting to move toward 2.5D and 3D packaging, but for the foreseeable future thermal limitations will prevent a fully 3D processor from being viable. Transistors per unit volume doesn't really have advantages over transistors per unit area when we're at most looking at stacking some cache on top of the CPU logic (and in the near future, we're really just putting chiplets alongside each other and providing interconnections with similar performance to communication across a large monolithic die).
    - titzer 5y ago
      You're totally right that heat is the enemy to 3D designs. They will only work with low power, or if nanoscale heat pipes can be integrated into the design.
- bryanlarsen 5y ago
  This should be upvoted to the top of the discussion. The article never makes a suggestion for what should replace nanometers as the node name. This is a great suggestion: million transisters per square millimeter.
- mjevans 5y ago
  That's a good start. Different applications would probably use different scores, based on a table of numbers. +++ Count of transistors within a single layer (E.G. what's linked above) Count of transistors within a 3d volume Performance (flops? some other standard?) of that volume under various thermal conditions, power constraints, etc. Latency at (the pins and) edges during the common test pattern (during the above profile slots). Efficiency: how much power is consumed in the above states. Weight; sometimes it matters, this should be measured and published even if it isn't used in the score.
Guthur 5y ago
No one beyond chip nerds like myself ever really hear or care about node stepping and we all know it's more of a name than anything meaningful. Intel don't market the node itself to consumers but rather the architectural generation. The whole 10 vs 7nm does not generate a perception of Intel being behind but rather they now are behind with the move to TSMC 5nm and the fact Intel still can't get volume production from 10nm for any product line that requires top performance. The cynic in me feels that this is recent FUD originated in Intel marketing to try and smoke screen the disaster that is Intel 10nm. A little too late imo. It's not that 10nn is bad but rather how long they've actually been working on it. If it had only taken twice as long to get to volume production they would have been well a ahead of the competition, remember the old tick/tock. Arguably 10nm was suppose to be here in 2016, but only actually went into production in 2018, hence the the 14nm+++, and it's still not there yet.
- Guthur 5y ago
  As an aside I think Intel have actually done phenomenal work to stretch their 14nm as far as they have. It's taken a significant amount of power draw to do it, but still a significant feat of semi conductor design imo.
  - techrat 5y ago
    Iterative designs for multiple generations with minimal IPC improvements across multiple sockets is "phenomenal work"? With all of the processor flaw mitigations, we saw the Lake series effectively lose all of their gains, however slim, between each generation. Rocket Lake, the newest, still gets a 2 to 16% performance hit (depending on task) with mitigations on. https://www.phoronix.com/scan.php?page=article&item=spectre-rocket-lake&num=7 https://www.phoronix.com/scan.php?page=article&item=spectre-... At this point, Intel is behind where they started with the Core line.
    - Guthur 5y ago
      They stretched the app core clocks quite a bit. For sure they had a lot of historical performance lead they had to burn to stay relevant but I don't think it would have been easy to get 14nm as far as they have. It has definitely got very little to give going forward though.
analog31 5y ago
In my view, there might be uses for different kinds of units, depending on what you care about. Nanometers are interesting to me because they relate to progress in lithography technology. Nodes or gates per acre tells us something. Achievable information storage or bandwidth per acre tells us something else.
Havoc 5y ago
Aka the numbers are made up. Not convinced it’s a problem though. Intel seems to be doing fine with their farcical 14+++ naming
- askvictor 5y ago
  Kind of the same as clock speed, though Intel was on the other side of the coin a decade or two ago, pushing the idea that clock speed is the only performance measure that matters (until it changed it's mind with the Core architecture).
dane-pgp 5y ago
(2020)
elihu 5y ago
Perhaps we should migrate to a "6502s per square milimeter" metric or something?
- raverbashing 5y ago
  No kidding, I think from some Larabee conversations that we're already at > 1 Pentium core per mm2 (the original P5)
SoylentOrange 5y ago
As a computer scientist rather than an EE, I learned a lot from this article. I would like to provide some perspective as a person who buys and looks at processor advances in the consumer market with a technical, but not too technical, background. Unlike a sibling comment, I never knew that the lithography process did not correspond to any physical characteristic in the chip. I naively assumed that the headlines saying that Intel was behind in process technology were correct, and used that to inform purchasing decisions. I found the article enlightening, and I think many on HN with the same background as myself will agree. As an aside, I tried to read the linked IEEE paper [1] but the page is cut off for me below section 2. If anyone has a link to the full PDF, I would appreciate that. [1]: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9063714 https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=906...
- wtallis 5y ago
  > I naively assumed that the headlines saying that Intel was behind in process technology were correct, and used that to inform purchasing decisions. Those headlines are correct, and were at the time this article was written as well. Intel's first 10nm process was completely broken, and their current less-dense 10nm process is still barely usable. Any density advantages Intel was planning on their 10nm having over TSMC 7nm are meaningless at this point. Higher density is pointless unless you can get yields that are good enough to ship chips that are meaningfully better on power, performance and price.
  - lorenzhs 5y ago
    It's not "barely usable" any more, they're shipping Xeons with up to 40 cores at volume. Afaik everything except high-clock desktop parts is shipping on 10nm now. It took them an eternity, but it's not Duke Nukem late.
    - wtallis 5y ago
      Those Xeons only make sense if you're able to make very good use of AVX-512 or if you're bureaucratically prevented from buying AMD. They aren't a disaster to the extent that Cannonlake was, but it's still a product that has trouble standing on its merits and is shipping in part because Intel couldn't have cancelled it and told customers to wait for Sapphire Rapids without facing a shareholder lawsuit for lying about the viability of their roadmap.
jabl 5y ago
Related article (that was previously discussed on HN): https://spectrum.ieee.org/semiconductors/devices/a-better-way-to-measure-progress-in-semiconductors https://spectrum.ieee.org/semiconductors/devices/a-better-wa... It covers the LMC proposal mentioned in this article, but also another proposal more similar to the current node naming, but instead of using the gate length, using a combination of gate pitch (minimum distance between two transistors) and metal pitch (minimum distance between two wires).
i6ruce 5y ago
I do not understand why would anyone care about nanometers at all. There are power efficiency, performance, reliability, price/value characteristics that are more important to an average consumer than the lithography process size.
- ur-whale 5y ago
  Because people like simple numbers. Examples: - megapixels for cameras - MHz for CPU's (back in the days) - HP for cars etc...
- tremon 5y ago
  Because the node size used to be a reliable proxy for power efficiency, performance and price characteristics. That was true for decades, until the node size became so small that the quantum effects changed the game.
mnw21cam 5y ago
So nanometres are the new megahertz.
salmo 5y ago
This whole thing seems like a wrong take. I don’t think TMSC even registers Intel as competition these days. Intel vs X is a tiny market. Most CPUs don’t go in x86 machines, and the chip companies choose fabs on so many more variables than just “nm”. Products pick chips on higher level stats than that: price, power, performance, features, footprint, heat, etc. Consumers usually don’t know or care what CPU is in their device (phones, tablets, watches, doorbells, coffee mugs, etc). They pick for features, fashion, and product ecosystem. And the cloud/data center market (maybe the last bastion of x86) only cares about price + performance per watt. I think the only people who pay attention to this stuff are PC gamers, and they’re a niche within a niche. And even that is more of a Ford vs Chevy market of brand loyalty. Maybe some market analysts use it to ding Intel, but I think that’s a minor detail in a much gloomier picture.

121 points by bcaa7f3a8bbc 5y ago

aappleby 5y ago

no mention of https://en.wikichip.org/wiki/mtr-mm%C2%B2 https://en.wikichip.org/wiki/mtr-mm%C2%B2 ?

Waterluvian 5y ago
I know very little about this, but shouldn't the ideal unit consider volume? They're going to eventually do clever 3d stuff I'm guessing.
- wtallis 5y ago
  CPUs are starting to move toward 2.5D and 3D packaging, but for the foreseeable future thermal limitations will prevent a fully 3D processor from being viable. Transistors per unit volume doesn't really have advantages over transistors per unit area when we're at most looking at stacking some cache on top of the CPU logic (and in the near future, we're really just putting chiplets alongside each other and providing interconnections with similar performance to communication across a large monolithic die).
  - titzer 5y ago
    You're totally right that heat is the enemy to 3D designs. They will only work with low power, or if nanoscale heat pipes can be integrated into the design.
bryanlarsen 5y ago
This should be upvoted to the top of the discussion. The article never makes a suggestion for what should replace nanometers as the node name. This is a great suggestion: million transisters per square millimeter.
mjevans 5y ago
That's a good start. Different applications would probably use different scores, based on a table of numbers. +++ Count of transistors within a single layer (E.G. what's linked above) Count of transistors within a 3d volume Performance (flops? some other standard?) of that volume under various thermal conditions, power constraints, etc. Latency at (the pins and) edges during the common test pattern (during the above profile slots). Efficiency: how much power is consumed in the above states. Weight; sometimes it matters, this should be measured and published even if it isn't used in the score.

Guthur 5y ago

No one beyond chip nerds like myself ever really hear or care about node stepping and we all know it's more of a name than anything meaningful. Intel don't market the node itself to consumers but rather the architectural generation. The whole 10 vs 7nm does not generate a perception of Intel being behind but rather they now are behind with the move to TSMC 5nm and the fact Intel still can't get volume production from 10nm for any product line that requires top performance. The cynic in me feels that this is recent FUD originated in Intel marketing to try and smoke screen the disaster that is Intel 10nm. A little too late imo. It's not that 10nn is bad but rather how long they've actually been working on it. If it had only taken twice as long to get to volume production they would have been well a ahead of the competition, remember the old tick/tock. Arguably 10nm was suppose to be here in 2016, but only actually went into production in 2018, hence the the 14nm+++, and it's still not there yet.

Guthur 5y ago
As an aside I think Intel have actually done phenomenal work to stretch their 14nm as far as they have. It's taken a significant amount of power draw to do it, but still a significant feat of semi conductor design imo.
- techrat 5y ago
  Iterative designs for multiple generations with minimal IPC improvements across multiple sockets is "phenomenal work"? With all of the processor flaw mitigations, we saw the Lake series effectively lose all of their gains, however slim, between each generation. Rocket Lake, the newest, still gets a 2 to 16% performance hit (depending on task) with mitigations on. https://www.phoronix.com/scan.php?page=article&item=spectre-rocket-lake&num=7 https://www.phoronix.com/scan.php?page=article&item=spectre-... At this point, Intel is behind where they started with the Core line.
  - Guthur 5y ago
    They stretched the app core clocks quite a bit. For sure they had a lot of historical performance lead they had to burn to stay relevant but I don't think it would have been easy to get 14nm as far as they have. It has definitely got very little to give going forward though.

analog31 5y ago

In my view, there might be uses for different kinds of units, depending on what you care about. Nanometers are interesting to me because they relate to progress in lithography technology. Nodes or gates per acre tells us something. Achievable information storage or bandwidth per acre tells us something else.

Havoc 5y ago

Aka the numbers are made up. Not convinced it’s a problem though. Intel seems to be doing fine with their farcical 14+++ naming

askvictor 5y ago
Kind of the same as clock speed, though Intel was on the other side of the coin a decade or two ago, pushing the idea that clock speed is the only performance measure that matters (until it changed it's mind with the Core architecture).

dane-pgp 5y ago

(2020)

elihu 5y ago

Perhaps we should migrate to a "6502s per square milimeter" metric or something?

raverbashing 5y ago
No kidding, I think from some Larabee conversations that we're already at > 1 Pentium core per mm2 (the original P5)

SoylentOrange 5y ago

As a computer scientist rather than an EE, I learned a lot from this article. I would like to provide some perspective as a person who buys and looks at processor advances in the consumer market with a technical, but not too technical, background. Unlike a sibling comment, I never knew that the lithography process did not correspond to any physical characteristic in the chip. I naively assumed that the headlines saying that Intel was behind in process technology were correct, and used that to inform purchasing decisions. I found the article enlightening, and I think many on HN with the same background as myself will agree. As an aside, I tried to read the linked IEEE paper [1] but the page is cut off for me below section 2. If anyone has a link to the full PDF, I would appreciate that. [1]: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9063714 https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=906...

wtallis 5y ago
> I naively assumed that the headlines saying that Intel was behind in process technology were correct, and used that to inform purchasing decisions. Those headlines are correct, and were at the time this article was written as well. Intel's first 10nm process was completely broken, and their current less-dense 10nm process is still barely usable. Any density advantages Intel was planning on their 10nm having over TSMC 7nm are meaningless at this point. Higher density is pointless unless you can get yields that are good enough to ship chips that are meaningfully better on power, performance and price.
- lorenzhs 5y ago
  It's not "barely usable" any more, they're shipping Xeons with up to 40 cores at volume. Afaik everything except high-clock desktop parts is shipping on 10nm now. It took them an eternity, but it's not Duke Nukem late.
  - wtallis 5y ago
    Those Xeons only make sense if you're able to make very good use of AVX-512 or if you're bureaucratically prevented from buying AMD. They aren't a disaster to the extent that Cannonlake was, but it's still a product that has trouble standing on its merits and is shipping in part because Intel couldn't have cancelled it and told customers to wait for Sapphire Rapids without facing a shareholder lawsuit for lying about the viability of their roadmap.

jabl 5y ago

Related article (that was previously discussed on HN): https://spectrum.ieee.org/semiconductors/devices/a-better-way-to-measure-progress-in-semiconductors https://spectrum.ieee.org/semiconductors/devices/a-better-wa... It covers the LMC proposal mentioned in this article, but also another proposal more similar to the current node naming, but instead of using the gate length, using a combination of gate pitch (minimum distance between two transistors) and metal pitch (minimum distance between two wires).

i6ruce 5y ago

I do not understand why would anyone care about nanometers at all. There are power efficiency, performance, reliability, price/value characteristics that are more important to an average consumer than the lithography process size.

ur-whale 5y ago
Because people like simple numbers. Examples: - megapixels for cameras - MHz for CPU's (back in the days) - HP for cars etc...
tremon 5y ago
Because the node size used to be a reliable proxy for power efficiency, performance and price characteristics. That was true for decades, until the node size became so small that the quantum effects changed the game.

mnw21cam 5y ago

So nanometres are the new megahertz.

salmo 5y ago

This whole thing seems like a wrong take. I don’t think TMSC even registers Intel as competition these days. Intel vs X is a tiny market. Most CPUs don’t go in x86 machines, and the chip companies choose fabs on so many more variables than just “nm”. Products pick chips on higher level stats than that: price, power, performance, features, footprint, heat, etc. Consumers usually don’t know or care what CPU is in their device (phones, tablets, watches, doorbells, coffee mugs, etc). They pick for features, fashion, and product ecosystem. And the cloud/data center market (maybe the last bastion of x86) only cares about price + performance per watt. I think the only people who pay attention to this stuff are PC gamers, and they’re a niche within a niche. And even that is more of a Ford vs Chevy market of brand loyalty. Maybe some market analysts use it to ding Intel, but I think that’s a minor detail in a much gloomier picture.