Tuesday, August 30, 2016

Techy Tuesday - Moore's Law Ended in 2012

That's the view of a column from Industry Observer Zvi Or-Bach in yesterday's EE Times online, and he makes a persuasive case.  In it, he points out two major trends: first, the cost per transistor stopped decreasing as semiconductor makers went through the 28 nanometer (28 nm) geometry in 2012.  (The size used to describe the semiconductor geometry is based on the smallest features in the part).  Today, the leading edge is 16nm, with Intel just announcing processors built to a 14nm process and talking about going to 7nm. 

What Or-Bach points out in the graph on the left is that we can still produce smaller transistors at the finest geometries available, but they aren't cheaper.  The graph shows the number of transistors (in millions per dollar) at a certain geometry, versus the year.  The climb (more transistors per dollar) until 2012 is obvious, as is the graph going flat, indicating a constant number per dollar, then fewer transistors per dollar as the geometry shrinks from 20 to 16 nm and the curve tilts down.  Since the early days of semiconductor production, each generation brought smaller and cheaper transistors.

The right side of the plot shows another interesting trend.  At 130 nm, there were 22 facilities that made integrated parts with lots of transistors ("fabs").  Today, only four fabs in the world are capable of working at 16 and 14 nm.  Going to smaller geometries entails more expenses: investment in both physical facilities; i.e, machines and buildings, as well as investment in the processes; learning how to actually make things that work.

The reality of the industry, though, is that these ultra-fine geometries are the glamor side of the business.  The actual work is done by older, less fine geometries. 
The graph on the left is showing that 43% of worldwide semiconductor production is in the five largest geometries: from 65nm up to the largest sizes used.  Further, the graph on the right shows that 85% of new designs are 65nm and larger.  Or-Bach takes this as evidence that the industry is bifurcating; very few designs move to the finest geometries while most designs are being designed into older, bigger geometries.  This is leading foundries to invest and develop enhancements to the older processes, keeping older facilities in production longer.

Saying that Moore's Law is over doesn't mean innovation is going to stop, but it does imply that the relentless trend of processors doubling in the number of transistors every two years is ending.  As we've noted before, processor speeds have been "stuck" at a few GHz for a decade or more.   We talked about 3D FLASH memory in this space back in January, and that's one example of innovations that increase density without going to finer geometries.  There are a lot of very clever engineers in that business, and I suspect we'll see more innovative approaches than just "transistor slinging" as it gets progressively less feasible to just throw more transistors at a design. 

Finally, just for fun, here's a "previous generation" processor from Intel.  It uses 22nm geometry. 
Note that it says there are 1.4 Billion transistors in that part.  I'm going to use Or-Bach's number of 20 million transistors per dollar, and predict this was a $70 part.  I don't know how many of these parts Intel made, or is making, but think of the number of transistors in these, then every other processor.  It's why I think that humanity has made more transistors (in particular, FETs) than any other thing we've made in all of human history.  More transistors than nails, sheets of paper, anything in history.


6 comments:

  1. Remember the good 'ol days when transistor radios had the "tin can" transistors with three legged leads, and a few extra "unused" cans thrown in just so the vendor could claim their device supported a bigger number of those amazing "transistor" devices? A portable radio with seven, wow, seven transistors!!

    ReplyDelete
    Replies
    1. Yes, I do. I think I had one.

      That was by the mid-60s, I'm thinking '64 or '65. Not too long after that, I'd take apart vacuum tubes to see how they were made. Tried that with one of those little metal can transistors. I think it was filled with silicon thermal grease.

      Delete
    2. If you hooked one of those old can-style transistors directly to a 9v battery, it would sometimes get hot enough to explode.

      What can i say. I was 12 at the time.

      Delete
  2. Do you have or know where a plot of $/transistor vs node size for present day is?
    While it's reasonable that the bleeding edge $/transistor has stabilized, or at least slowed (given the number of multiple exposures, etc to get sub-sub-sub-wavelength features and associated process steps and masks), the larger should be getting cheaper faster.

    Basically, the production rate vs node size plots you have, but $/transistor vs node size.

    Thoughts?

    ReplyDelete
    Replies
    1. Very interesting question - unfortunately, I don't know where to find such a plot. As you say, I'd expect the $/transistor is probably still going down in the older processes.

      The column that launched this piece was written by Zvi Or-Bach and it says he's the founder of MonolithIC 3D Inc. I thought they were a consulting firm, but they are (as the name implies) a company that developed a process for 3D ICs. My guess is that one of the companies that tracks the semiconductor industry probably has a chart like you're looking for. I'm going to look a bit and see if I can find such a thing.

      Delete
    2. I agree, it should be dropping due to a few things: 1. Previous node equipment available at low cost to make room for new, 2. Improved handling/alignment/etc tech improving yield of older processes, 3. Ability to used triaged (i.e. Didn't meet QC for the most advanced node) materials/etc at lower cost. 4. Ability to take advantage of mask cost reductions--since most fab is UV/EUV, and thus -any- node smaller than 90-100nm is going to be a multiple-mask process (really small nodes requiring a relatively large number of masks), the cost per mask is driven down, so older nodes using fewer masks can take advantage (since they are likely all the same physical size)

      If you find something, let us know, I'll look as well and do the same.

      Delete