Why 45nm?

[Pederv]

Is it the math or did somebody pull the number out of their a$$?

 

[Diogenes2]

It probably has something to do with using the same size silicon wafers as their 90nm fab ..
Lets them continue to use a lot of the same machinery ..

 

[tigersty1e]

Because when you process the chips, 45nm offers cheaper technology and more possible improvements in the future. You can't go all out in the first one.

 

[Miramonti]

Originally posted by: Diogenes2
It probably has something to do with using the same size silicon wafers as their 90nm fab ..
Lets them continue to use a lot of the same machinery ..

Then why 90nm...and then 65nm?

I'm curious too. There must be some logical reason they are all divisible by 5.

 

[AmberClad]

Originally posted by: jjsole
There must be some logical reason they are all divisible by 5.

Apparently, they're not always divisible by 5. The ones so far have been, but the chart seems to indicate that future processes will not be.

 

[ArchAngel777]

It is probably a percentage of sorts, not to mention a limitation with the technology itself.

 

[pm]

The process technology "nodes" are determined by the International Technology Roadmap for Semiconductors (ITRS) working group.
http://www.itrs.net/reports.html

Basically a large world-wide group decides which process technology nodes are on the "roadmap" and they standardize the manufacturing toolset around the technology requirements that they set. So they decided literally over a decade ago that the industry would be using 45nm and then highlighted the requirements to achieve this, what areas were of concern and help drive the solutions. The advantage of this approach is that manufacturing tools then are more interoperable than they otherwise might be, and there's a global concerted effort to standardize on a given process technology "node".

As to how these numbers are chosen, it's a percentage reduction coupled with a bit of rounding. And there are more "nodes" than the big publicized ones of 90nm, 65nm and 45nm.

The 2001 ITRS roadmap for MPUs (microprocessors) has:
130nm 2002
107nm 2003
90nm 2004
80nm 2005
70nm 2006
65nm 2007

and the 2006 issue has:
78nm 2006
68nm 2007
59nm 2008
52nm 2009
45nm 2010

(follow the roadmap link above, click on lithography and then go to about page 3-6 (Technology Requirements)

 

[Pederv]

Thanks PM

So it is possible to "fine tune" equipment designed for 45nm to create nodes that are 38nm (which is halfway between 45 and 32)?

I was just wondering why where Intel is going from 65nm to 45, to 32, to whatever is next over the next 6 years, AMD couldn't go from 65nm to 38, to whatever is next over the same 6 years. Maybe costing AMD less to maintain a relatively close node size.

 

[dandragonrage]

90/sqrt(2) = 65
65/sqrt(2) = 45
45/sqrt(2) = 32
and so on

 

[XBoxLPU]

Originally posted by: Pederv

I was just wondering why where Intel is going from 65nm to 45, to 32, to whatever is next over the next 6 years, AMD couldn't go from 65nm to 38, to whatever is next over the same 6 years. Maybe costing AMD less to maintain a relatively close node size.

$$$$

 

[996GT2]

Originally posted by: dandragonrage
90/sqrt(2) = 65
65/sqrt(2) = 45
45/sqrt(2) = 32
and so on

Actually, 90/sqrt(2) is 63.64, though the others are pretty close to 65 and 45. Do Intel and AMD really just round like that?

 

[JAG87]

Originally posted by: 996GT2

Originally posted by: dandragonrage
90/sqrt(2) = 65
65/sqrt(2) = 45
45/sqrt(2) = 32
and so on

Actually, 90/sqrt(2) is 63.64, though the others are pretty close to 65 and 45. Do Intel and AMD really just round like that?

He's close but hes wrong. Its second last lithography process divided by 2.

180nm/2 = 90nm
130nm/2 = 65nm
90nm /2 = 45nm
65nm /2 = 32nm
45nm /2 = 22nm
32nm /2 = 16nm

and like Diogenes said, it does have to do with wafer size.

 

[21stHermit]

Originally posted by: pm
The process technology "nodes" are determined by the International Technology Roadmap for Semiconductors (ITRS) working group.
http://www.itrs.net/reports.html

Basically a large world-wide group decides which process technology nodes are on the "roadmap" and they standardize the manufacturing toolset around the technology requirements that they set. So they decided literally over a decade ago that the industry would be using 45nm and then highlighted the requirements to achieve this, what areas were of concern and help drive the solutions. The advantage of this approach is that manufacturing tools then are more interoperable than they otherwise might be, and there's a global concerted effort to standardize on a given process technology "node".

As to how these numbers are chosen, it's a percentage reduction coupled with a bit of rounding. And there are more "nodes" than the big publicized ones of 90nm, 65nm and 45nm.

One of the more informative post I've seen on this or any other forum!!! :beer: :beer:

 

[CTho9305]

Originally posted by: Pederv
So it is possible to "fine tune" equipment designed for 45nm to create nodes that are 38nm (which is halfway between 45 and 32)?

If you look back a TSMC's manufacturing nodes you'll see that they've done a lot of "half-nodes" (e.g 110nm, 80nm, 55nm) , which have shown up in various GPUs.

I haven't really thought this idea through, but I'll post it anyway: it makes sense for ASIC foundries to ship half-nodes because the design effort to remap a design is much lower. Shrinking a modern processor (e.g. 90nm -> 65nm) takes a full design team a good year; even though a 90->80nm shrink would probably be easier, it would still be a huge task. It's not just design effort: shipping half-node products would presumably double the number of silicon variants, thus doubling the resources required from the stages between tape out and release (I'll use the term "validation" broadly here; this generally takes about a year). There are probably also yield learning curves that would also come into play in the actual manufacturing.

ASIC designs like GPUs or network controllers tends to be quicker to port between processes, since they're often almost completely designed using automated synthesis, placement and routing; re-spinning a chip with a new library takes very little time.

I was just wondering why where Intel is going from 65nm to 45, to 32, to whatever is next over the next 6 years, AMD couldn't go from 65nm to 38, to whatever is next over the same 6 years. Maybe costing AMD less to maintain a relatively close node size.

The basic trade-off is up-front R&D cost versus per-die cost. Using smaller manufacturing processes lets you fit more dies on a wafer, lowering the per-die cost. Bringing up a new process node takes a huge amount of work and money. If you decide to skip a node, you'll be suffering when your chips cost you 50% more to manufacture than the competitor's during the time when they've switched over and you haven't. If you try to move forward more aggressively to new nodes, you have to do a lot more of the learning yourself and can't leverage work done by other companies.