Why not skip some process nodes and jump ahead e.g. to 5 nm directly?

[Fjodor2001]

Hi,

I just wonder how come the process node advancements have to pass through all steps, i.e. 22->14->10->7->5->[...] nm?

Wouldn't it be an option for some semiconductor manufacturer to jump ahead, and e.g. skip 14, 10 and 7 nm and go directly to 5 nm? Sure it would take more time to go from 22->5nm than going from e.g. 22->14 nm. But couldn't it be quicker than going through each of those individual process nodes (22->14->10->7->5nm)?

Or is there some physical logic behind this decision. For example that in order to produce 10 nm chips, you must also have solved the previous problems related to 14 nm anyway?

I'd be interested in hearing some opinions on this. :hmm:

 

[jhu]

Because those number don't really mean much anymore. They could call the next node "secret magic pixie fairy dust", and it would still have the same meaning.

 

[Fjodor2001]

Because those number don't really mean much anymore. They could call the next node "secret magic pixie fairy dust", and it would still have the same meaning.

Yes, but you still get the point, right? Regardless if the process nodes are numbered 14 or 16 nm, why not skip ahead some nodes?

 

[MagnusTheBrewer]

Hi,

I just wonder how come the process node advancements have to pass through all steps, i.e. 22->14->10->7->5->[...] nm?

Wouldn't it be an option for some semiconductor manufacturer to jump ahead, and e.g. skip 14, 10 and 7 nm and go directly to 5 nm? Sure it would take more time to go from 22->5nm than going from e.g. 22->14 nm. But couldn't it be quicker than going through each of those individual process nodes (22->14->10->7->5nm)?

Or is there some physical logic behind this decision. For example that in order to produce 10 nm chips, you must also have solved the previous problems related to 14 nm anyway?

I'd be interested in hearing some opinions on this. :hmm:

You answered your own question. The manufacturing process has to be refined a step at a time. It would take longer and cost more to attempt to "skip " major process nodes. You would introduce so many new variables, you wouldn't know what to fix or fix first.

 

[daxzy]

The numbers are developed by a standardized organization. They're approximate halves every other generation.

250 nm – 1997
180 nm – 1999
130 nm – 2002 (~1/2 250)
90 nm – 2004 (1/2 180)
65 nm – 2006 (1/2 130)
45 nm – 2008 (1/2 90)
32 nm – 2010 (~1/2 65)
22 nm – 2012 (~1/2 45)
14 nm – 2014 (this one's a bit off from 1/2 32)
10 nm – est. 2015 (~1/2 22)
7 nm – est. 2017 (1/2 14)
5 nm – est. 2019 (1/2 10)

 

[witeken]

Maybe that could have been possible some decades ago, but I think today, if you can follow Moore's law, you're already doing very good.

There is a lot of R&D necessary for all those nodes apart, so it's going to be very difficult if you want to do everything in 1 time. You can better just do it a bit "more slowly". Image if you fail and it takes a lot longer than expected, then you have just an outdated process node. So the risk/reward is simply too high. It also doesn't make sense from an economical standpoint. Following Moore's law already gives very nice benefits, so why would you want to do it even faster?

Also, when you ask a company like Intel, they'll say that they have a clear visibility ~5 years ahead, so if you're going to skip 3 nodes (~6 years), you don't really know what kind of problems you're going to get. Advancing nodes simply takes time.

Finally, no company does everything. TSMC, Intel, Samsung, Global Foundries,... all depend on other companies like AMSL and Applied Materials. They all have roadmaps too. So if you want to skip some nodes to get a massive lead on the competition, you have to convince them (probably with money) to develop the required tools faster. And you have to take into account that the competition then will also be able to go to that same node, or you'll have to do it on your own, which also costs a lot of money.

International Technology Roadmap for Semiconductors

Conclusion: it's unlikely that nodes will be skipped to advance faster. Following Moore's law is already difficult enough.

 

[witeken]

Because those number don't really mean much anymore. They could call the next node "secret magic pixie fairy dust", and it would still have the same meaning.

A new node does have, to some extend, a real meaning. Simply put, a new node should give roughly double the amount of transistors/area. just like clock speeds, it's also not easy to use them to see who has the best node. TSMC is going to break with the node naming even more at 16nm, which is as far as I know the first node that doesn't give about a doubling in density.

 

[Lonyo]

Yes, but you still get the point, right? Regardless if the process nodes are numbered 14 or 16 nm, why not skip ahead some nodes?

The next "node" is simply what's expected to be achievable after the current one. If it was possible to skip ahead, pretty much everyone would do it and instead of being a "skip", it would simply be just the next node.

The reason it's thought to be the most achievable is because it gives enough of an increase in density without being too complex to take the step required to get there (e.g. refining existing processes).
Some steps do require more effort with certain areas of the process, which slows them down, but in those cases it's easier to start "big" before trying to go tiny with brand new production processes.

If you could skip ahead, then simply put there wouldn't be those intermediate nodes, because they wouldn't be "needed".

 

[witeken]

The numbers are developed by a standardized organization. They're approximate halves every other generation.

250 nm – 1997
180 nm – 1999
130 nm – 2002 (~1/2 250)
90 nm – 2004 (1/2 180)
65 nm – 2006 (1/2 130)
45 nm – 2008 (1/2 90)
32 nm – 2010 (~1/2 65)
22 nm – 2012 (~1/2 45)
14 nm – 2014 (this one's a bit off from 1/2 32)
10 nm – est. 2015 (~1/2 22)
7 nm – est. 2017 (1/2 14)
5 nm – est. 2019 (1/2 10)

The problem becomes already visible with the official node names. 14nm used to be 16nm, and also 7nm is/was sometimes referred to as 8nm.

 

[sefsefsefsef]

The numbers are developed by a standardized organization. They're approximate halves every other generation.

250 nm – 1997
180 nm – 1999
130 nm – 2002 (~1/2 250)
90 nm – 2004 (1/2 180)
65 nm – 2006 (1/2 130)
45 nm – 2008 (1/2 90)
32 nm – 2010 (~1/2 65)
22 nm – 2012 (~1/2 45)
14 nm – 2014 (this one's a bit off from 1/2 32)
10 nm – est. 2015 (~1/2 22)
7 nm – est. 2017 (1/2 14)
5 nm – est. 2019 (1/2 10)

The sequence of nodes actually is given by the formula: f(x+1) = f(x)/1.44

They reduce the x and y measurement of features by sqrt(2), so that each feature takes up 1/2 the area of the previous generation, so you get 2x more transistors in the same area.

The reason they don't skip generations is economics. They take baby steps because they believe they'll get a good return on investment. If in 2010 Intel had wanted to jump straight to the 5nm process, it probably would have still taken them until 2019 to pull it off, but they would not be making any money in the intervening years.

 

[videogames101]

because each node is as advanced as possible at the time - if a node could be "skipped" to it would instead just be the next node

 

[KingFatty]

Why eat breakfast, lunch, and dinner each day? Just wake up and eat dinner, think of all the time saved.

 

[Fjodor2001]

Why eat breakfast, lunch, and dinner each day? Just wake up and eat dinner, think of all the time saved.

Why eat breakfast, when you could split it up into early-breakfast, mid-breakfast, and late-breakfast?

 

[Boze]

Why eat breakfast, when you could split it up into early-breakfast, mid-breakfast, and late-breakfast?

Because you'd still have to eat second breakfast, of course!

 

[Fox5]

They probably could skip a node or two, at exponentially increasing costs. It takes a while before yields on a node are good enough to get a ROI.

 

[OBLAMA2009]

Hi,

I just wonder how come the process node advancements have to pass through all steps, i.e. 22->14->10->7->5->[...] nm?

Wouldn't it be an option for some semiconductor manufacturer to jump ahead, and e.g. skip 14, 10 and 7 nm and go directly to 5 nm? Sure it would take more time to go from 22->5nm than going from e.g. 22->14 nm. But couldn't it be quicker than going through each of those individual process nodes (22->14->10->7->5nm)?

Or is there some physical logic behind this decision. For example that in order to produce 10 nm chips, you must also have solved the previous problems related to 14 nm anyway?

I'd be interested in hearing some opinions on this. :hmm:

why would a business do something that stupid, so we could all have computers that would be good for the next 50 years?

 

[tweakboy]

Hi,

I just wonder how come the process node advancements have to pass through all steps, i.e. 22->14->10->7->5->[...] nm?

Wouldn't it be an option for some semiconductor manufacturer to jump ahead, and e.g. skip 14, 10 and 7 nm and go directly to 5 nm? Sure it would take more time to go from 22->5nm than going from e.g. 22->14 nm. But couldn't it be quicker than going through each of those individual process nodes (22->14->10->7->5nm)?

Or is there some physical logic behind this decision. For example that in order to produce 10 nm chips, you must also have solved the previous problems related to 14 nm anyway?

I'd be interested in hearing some opinions on this. :hmm:

It's called MONEY!!! :whiste:

 

[mrmt]

Or is there some physical logic behind this decision. For example that in order to produce 10 nm chips, you must also have solved the previous problems related to 14 nm anyway?

There's an economic logic. Someone would have to first pay the R&D for the 5nm tools alone, which don't exist right now (ASML is shipping 10nm tools now). Also 5nm will need different engineering solutions.

Just look what happened from 65nm to 14nm, at 45nm we got HKMG, at 22nm we got finfet, and these solutions will be more or less mature at 10nm. 10nm and 7nm will need a new set of solutions but it will leverage on R&D from previous nodes. By going straight to 5nm someone would need to leverage on a lot of immature engineering solutions, which would bring heavy risks to the R&D project, both in costs and in time.

 

[norseamd]

why would a business do something that stupid, so we could all have computers that would be good for the next 50 years

just think of how good the current technology is

if we made all ip public domain we could double or even more the performance of processors. then all ip could be used on a single chip. would be nice to use some of those chips that certain hpc use

http://en.wikipedia.org/wiki/POWER8

how about one of this with a gpu for sending out video and a hpc class compute gpu for doing various gpgpu tasks. add a sound card and a wireless card and you have a amazing gaming computer

 

[carop]

Hi,

I just wonder how come the process node advancements have to pass through all steps, i.e. 22->14->10->7->5->[...] nm?

Wouldn't it be an option for some semiconductor manufacturer to jump ahead, and e.g. skip 14, 10 and 7 nm and go directly to 5 nm? Sure it would take more time to go from 22->5nm than going from e.g. 22->14 nm. But couldn't it be quicker than going through each of those individual process nodes (22->14->10->7->5nm)?

Or is there some physical logic behind this decision. For example that in order to produce 10 nm chips, you must also have solved the previous problems related to 14 nm anyway?

I'd be interested in hearing some opinions on this. :hmm:

FinFET silicon will be used for 14nm and 10nm devices. However, the manufacturers will have to use alternative channel materials for sub-10nm devices, and there are a number of unknowns regarding post-silicon devices.

The following article series cover the alternative channel materials for sub-10nm devices:

http://semiengineering.com/alternative-channel-materials-post-silicon-finfets-selective-epitaxy/

http://semiengineering.com/whats-after-silicon/

http://semiengineering.com/the-growing-list-of-unknowns-after-cmos/

 

[Arkadrel]

Why not skip some process nodes and jump ahead e.g. to 5 nm directly?

Because R&D would be insane (likelyhood of failour would be sky high, would be a huge gamble).

bascially what was said before, but said better than I could:

The manufacturing process has to be refined a step at a time. It would take longer and cost more to attempt to "skip " major process nodes. You would introduce so many new variables, you wouldn't know what to fix or fix first.

^ its just a easier, faster, cheaper way to do things, with higher returns on R&D.
Its also alot safer this way (money wise).

And whos to say that even if you tried to skip over alot of steps, that you would actually get someplace faster than those that took many small steps to reach that point?

 

[OBLAMA2009]

just think of how good the current technology is

if we made all ip public domain we could double or even more the performance of processors. then all ip could be used on a single chip. would be nice to use some of those chips that certain hpc use

http://en.wikipedia.org/wiki/POWER8

how about one of this with a gpu for sending out video and a hpc class compute gpu for doing various gpgpu tasks. add a sound card and a wireless card and you have a amazing gaming computer

is that going to be used in pc's and is it really going to be out this year?

 

[norseamd]

is that going to be used in pc's and is it really going to be out this year?

not sure what it will be used in. probably not any gaming pcs unless some company decides to try some thing different.

how does this power 8 compare with intel

 

[UaVaj]

5nm already available and locked up somewhere.

why bring out 5nm when you can continue to milk 22nm? only a fool would bring out 5nm now.

 

[Wall Street]

As others have said, they would need to change everything. Improved light sources, better patterning, new materials, lower leakage. When they progress one node at a time, they can experiment with one or two elements and figure out how changing those items impacts the overall process. It has been quite some time since a node has been canceled, but it has happened in the past where one of the foundries got too aggressive and realized half way through development that their experimental process would not work and had to go back to the drawing board. Given the billions that each node costs in R&D, if you spend the price to develop a node and end up with something that can't produce shipping products with good yield it can put your fab into major financial trouble.

A good example is extreme ultra violet light sources. For the past five years or so, a lot of people have thought that this was the future. However, to this day, nobody has been able to develop a powerful enough EUV light. Therefore, if you invested a lot of money in a process and EUV was on the extended list of things you MUST have for your investment to work, then you machines are sitting idle. The fabs want to avoid this situation, so they only make investments in new nodes when they see a clear near-term solution for all of the engineering hurdles.