backtop


Print 57 comment(s) - last by JumpingJack.. on Oct 12 at 10:50 PM


A young Gordon Moore circa 1975  (Source: Reuters)
Gordon Moore took part in the afternoon keynote during IDF 2007

The fall edition of the 2007 Intel Developer Forum (IDF) is officially underway from the San Francisco Moscone Center.  The Tuesday morning keynote featured more details about the Nehalem architecture as one of the main points of the discussion.

In a later session, Dr. Moira Gunn, host of NPR Tech Nation, hosted a fireside chat with Gordon Moore, Intel co-founder and creator of Moore's Law.  Moore received a well-deserved standing ovation from the crowded conference hall packed with thousands of attendees more than willing to respect a Silicon Valley legend.  

Of course, the question on everyone's mind was the validity of Moore's Law. Specifically, whether or not it holds up today the same way it did when Moore first documented his observations almost forty years ago.

Moore's Law -- actually more of a conjecture -- essentially states the number of transistors placed on an integrated circuit doubles every two years.  His observation helped outline trends the semiconductor industry for more than 40 years. 

"We have another decade, a decade and a half, before we hit something that is fairly fundamental," Moore said during the session.  That something "fundamental" is material science.  Even the most advanced lithography conceivable today can't eliminate the brick wall that is the nanoscale. 

Even at some point, lining up individual atoms no longer becomes feasible for transistor design.  Researchers from Intel are already easing into the field of using carbon nanotubes for processor interconnects; a team from the University of Pennsylvania just announced a new method for storing data via phase-changing nanowires.

"It's an exciting time," he said.  "I'd love to come back in 100 years and see what happened in the meantime."

Of course, even Moore's understanding of transistor trends is no match for the prowess of ambitious engineers. Conventional computing principles go out the window with the advent of quantum computing, for example.  Other types of alternative computing, including biological-based neural-computing, does not readily translate to transistor-count -- but that hasn't stopped researchers from making enormous progress in the last few years.

The death of Moore's Law is imminent, but new research and new materials assure that its successor will pack the same punch.


Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Already dead
By masher2 (blog) on 9/20/2007 8:59:17 AM , Rating: 2
> "Moore did nothing other than put a time scale to this concept..."

Of course, thats exactly what I said. Moore simply mated a statement of geometry to the the assumption that new process nodes would appear in linear time.

> "...postulating that it would take roughly 2 years to develop, debug, and put into production a new shrink"

But here's where the problem comes in. Moore didn't postulate "roughly" two years. He postulated ONE year. Intel, realizing the marketing value of "Moore's Law" has since restated it to 18 months, then 2 years. Soon it'll be 3 years.

Why? Because we don't move to new process nodes in linear time. Subsequent nodes become progressively harder. In 1970, we could move to new nodes each year. At 22nm at beyond, it will most likely be 3+ years.


RE: Already dead
By JumpingJack on 9/20/2007 10:19:24 PM , Rating: 2
quote:
Of course, thats exactly what I said. Moore simply mated a statement of geometry to the the assumption that new process nodes would appear in linear time.


No that is not what you said... what you said was:

quote:
In my opinion, Moore's Law already died quite some time ago.


This in simply not true.... each time we see a new node appear 2 years after the prior, Moore's law is being followed.

quote:
But here's where the problem comes in. Moore didn't postulate "roughly" two years. He postulated ONE year. Intel, realizing the marketing value of "Moore's Law" has since restated it to 18 months, then 2 years. Soon it'll be 3 years.


This is again untrue. Gordon Moore's first paper in 1965 was an invited paper, basically asking him to predict the state of the microelectronics industry over the next 10 years (in 1965), as you state he predicted a doubling ever year.
http://download.intel.com/museum/Moores_Law/Articl...

However, he came back, again per request, and revisted his predictions and restated what we know today as Moore's Law.
http://download.intel.com/museum/Moores_Law/Articl...

First he plotted DENSITY as a fuction of time:
quote:
Density can be expected to be proportional to the reciprocal of area, so the contribution to improve density vs. time from the use of smaller dimensions is plotted in Figure 3.


and concluded:
quote:
The new slope might approximate a doubling every two years, rather than every year, by the end of the decade.


So, it is pretty clear, transistor density (the correct form of Moore's law) will double ever two years. That is what is stated. I have posted above to show you that each node halves the area thereby doubling the density over the prior node... therefore, if a one node follows the prior by 2 years, Moore's Law holds.

Gordon Moore did not make up his law... he simply stated an observation on the trend in the data, which holds true today. He attached a time scale to what was, at the time, a standardize method of shrinking -- i.e. scale with a factor that will double the denisty.

In 1995, Gordon Moore wrote another analysis to see how well his prediction held up:
http://download.intel.com/museum/Moores_Law/Articl...

Pretty good so far.... what about today..

Well, 2001 Intel ships 130 nm... 2003 Intel ships 90 nm... 2005 Intel ships 65 nm and in 2007 they will ship 45 nm, and according to you (and Intel) they will ship 32 nm in 2009.... so it looks like Moore's Law is safe, at least until 2009.
http://download.intel.com/technology/silicon/techl... (see page 13, looks like every two years to me).

Now... if you want to see a linear plot, one of the best ways is to look at the reported 6T SRAM cell size over time, as SRAM is the highest density, most tightly packed array of 6 transistors -- in fact, this is on of a handful of parameters companies announce to tout their accomplishement.

See page 29 (same link as just above). So as you say, it should be exponential -- thus a logarithmic y-axis plotted against a linear axis of time (x-axis) should yield a straight line... well, it looks pretty straight to me.. all the way to 45 nm.

Does this make you wrong... not necessarily, it makes you wrong in your interpretation of Moore's Law... but to see if Moore's law really dies ... we will need to see what happens 2 years from about now -- as 45 nm launching on Nov 12th not only demonstrates Moore's Law it validates as it is about 2 years from Intel's 65 nm Process (actually a little less, but whose counting).

Where you are right is that there are other features that scale within the transistor that have reach physical limits... example, the gate dielectric material. In the transition from 90 nm to 65 nm, for the first time in history, the gate oxide thickness did not scale. Both Intel and AMD had a 1.2 nm gate oxide thickness at 90 nm and 1.2 nm gate oxide thickness at 65 nm.... why? The phenomena is called quantum mechanical tunneling. A more concise model base on QMT is called Fowler-Nordhiem Tunneling (google it, very educational).

However, scaling limits within the transistor aside... this does not invalidate Moore's Law... Moore's Law is a econmical cost structure model and only concerns itself with transistor density and cost of manufacturing. In fact, his original plot leading to this law in 1965, that famous hand sketched diagram, was a cost model not an engineering model.

Fascinating, and if you are further interesting in learning more about scaling theory in semiconductors, IBM has published a few very interesting articles as such:
http://www.research.ibm.com/journal/rd/462/taur.pd...
http://www.research.ibm.com/journal/rd/462/frank.p...

Jack


RE: Already dead
By flipsu5 on 9/20/2007 10:54:51 PM , Rating: 2
JJ, I think you take the node numbers too seriously. The design rules on CPUs are nowhere near those numbers. The pitch (gate-to-gate) needs to be really loose if you strain silicon to get fast chips (usually ~200 nm). On the other hand, memory doesn't care so much about speed but more so about density. That is where the transistor density scaling really follows the node numbers. The leader there is Samsung with 38 nm NAND flash. Toshiba is proposing to beat that with 3D array of transistors.


RE: Already dead
By JumpingJack on 9/20/2007 11:52:11 PM , Rating: 2
Well, not really ... The node, as defined fixes the pitch between the first two metal lines of the first metallization layer. ITRS defines the process node in their roadmap:
http://www.itrs.net/Links/2005ITRS/ExecSum2005.pdf page 6 has a nice diagram and the definition above..

People often think that the litho node refers to the smallest feature size on the transistor, this is not true. The smallest feature size that must be patterend is the Lg, (*distance between source and drain, making up the channel*) and depending on design, will be roughly 1/2 to 1/3 of the distance 1/2 Pitch at M1 -- at least that is the trend I have noted.

So at 65 nm, roughly 30 nm or so... in fact, both AMD and Intel have published 35 nm gate lengths at the 65 nm node, so that range is close. The TCAD, which designs the transistor has to make the gate to gate pitch commensurate with the M1 to M1 metal line pitch, otherwise, you would not be able to wire up the transistor. Going from your description, say left edge of the gate to the same gate edge on the adjacent transistor defines the total area taken by the transistor (source, channel, drain) plus the isolation between... the total pitch relative to M1 (as two adjacent M1 lines span a souce and a gate, sould then be 2xnodex2 roughly, so at 65 nm, for example, 2 x 65 x 2 give 260 nm of space to build the transistor and isolate it, at 45 nm this goes down to 180 nm.

The key to CADing out the transistor is maximize the overall drive current at the lowest leakage. Now, you mention stress ... good. Because in the absense of classical scaling, Intel and AMD turned to stress engineering to increase mobility as opposed to using geometry to drive up drive currents. This works good at 90 nm, not as well at 65 nm and will yield diminishing returns the smaller you go.... it is not stress that is actually phenomena, it is strain which is induced by stress. Stress is a pressure, strain though is a force... pressure is force/unit area... so if I stress over say area A1 and in my next revision stress over area A2 such that A2 < A1, then the induces strain is significantly less in the shrink.

Interestingly, AMD employed 2 major stressors in their 90 nm process (dual stress liners), but going to 65 nm the effectiveness of this stressing technique diminished enough that they obviously tried to put in two more (embedded SiGe and stress memorization), and even then they are struggling to recover and hit 90 nm speed bins... so it is tough, but it can be done even at 65 nm dimensions.

Jack


RE: Already dead
By JumpingJack on 9/21/2007 12:17:45 AM , Rating: 2
Let me add to your comment though, you bring up a good point. Not all transistors within a CMOS device are equal, some are made and patterned larger than others.... the definition does account by stating the to closest lines.

The point, I think you are making, is that we rarely see a perfect 50% compaction. This is true, so in the strictest sense there is some 'noise' about the line that makes up Moore's Law. I have seen Intel's compactions range from as high as 47% to as low as 43% from published data. AMD does not make that data readily available, however, the Brisbane product gives us an idea about 65 nm.... AMD only achieved about 31% compaction (not very good) for the same architecture and transistor count (two big stipulations you must make to draw anykind of corollary), which means something is wrong... their 65 nm is really more like 75 nm :) ....


RE: Already dead
By flipsu5 on 9/25/2007 8:35:22 AM , Rating: 2
It's true that metal one half-pitch is the main guide for ITRS but this is different for memory vs. other types of chips, e.g., ASICs or CPUs. 65 nm node Intel Metal 1 is 210 nm pitch (which would mean 105 nm design rule).It was 220 nm pitch for 90 nm node (or 110 nm design rule). The trick to shrinking the non-memory chips is to make them more memory-like.


RE: Already dead
By JumpingJack on 10/12/2007 10:50:02 PM , Rating: 2
This is incorrect... could you provide a linke to where you get this information, I think you are quoting gate to gate pitch, not metal to metal pitch.


"If they're going to pirate somebody, we want it to be us rather than somebody else." -- Microsoft Business Group President Jeff Raikes














botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki