With an IQ of 181, you might think it’s hard for me to amuse myself, but it’s not true. I like a little light reading in the evening…things like Kreck and Lück’s The Novikov Conjecture (A group G is a-T-menable, or, equivalently, has the Haagerup property if G admits a metrically proper isometric action on some affine Hilbert space, page 205) or Carver Mead’s Collective Electrodynamics (Because the system spatial attributes of the system in an eigenstate are stationary, the system in an eigenstate cannot radiate energy, page 106.)
Fascinating, riveting stuff, I’m sure you agree.
This led me to think a little bit about semiconductor reliability, though I confess my thinking might have been influenced by the consumption of a bottle of Puligny Montrachet Grand Cru Domaine des Comtes Lafon. In case you’re unfamiliar, that’s a wine very similar in appearance to the Franzia Crisp White wine you can buy in a box at the local market.
I mentioned Carver Mead above. Among his many accomplishments, he was the guy who contributed to Moore’s Law (which says transistor density on an integrated circuit will approximately double every two years) when Gordon Moore worked at Fairchild Semiconductor in the 1960s.
In 1965, Gordon had just started making his plots, where he’d plot the logarithm of the number of transistors on a chip as a function of the year. They’re little hand-drawn plots. I still have some around. One day we were talking about his plots. He said “You’re working on electron tunneling that happens when things get really small, right?” Yeah. “Well, wouldn’t that limit how small you can make a transistor?” Yes. “Well, how small is that?” Gordon has a way of asking these very simple questions that you really think you should know the answer to, and I didn’t. I said, well, I have to go and think about it. I’ve been thinking about it ever since. Carver Mead, from a speech at Telecosm 2006
The driver for Moore’s Law is the fact that as we make transistors smaller, they get cheaper to manufacture AND they work better. Isn’t that something? They work better. Think about that. And, what do we mean by working better? They consume less power and they can switch more quickly. This is the small miracle that fuels the marvelous advances of the digital revolution.
Today Fairchild is a leader in power management and mobile technology and in our factories, over 50 years after Fairchild Semiconductor was originally formed, we continue to take advantage of the state-of-the-art in transistor lithography.
There are not many businesses like ours where customers expect continuous improvement in products-with lower prices every year. Generally, if you buy the bargain brand at the big-box discount store, you expect lower quality…fewer features and less reliability. If you buy an inexpensive car, you expect it to be less comfortable and less reliable than a premium brand. But, that’s not how the semiconductor business works. Regardless of what you pay for our product, you expect high reliability…robust and rugged components.
Year-by-year, we practice continuous improvement and do our part to produce parts with very low failure rates. As a semi-random example, we publish a 3.65 FIT (Failure in Time) rating for the n-channel FET NDT3055. This works out to one failure for 3,127 years of operation.
Let’s take a closer look at what this means. We can’t build up a significantly significant number of parts and test them for 3,127 years. We’d love to, but we can’t. The FIT rating is based on accelerated testing of sample parts with numbers plugged into a formula.
The basic idea is that by aging sample devices with high humidity and overvoltage stress, we can estimate the failure rate without waiting for tens of thousands of years.
As I mentioned above, we do our part to create robust devices for our customers. However, is there also a role for the customer? Of course there is.
In building up our reliability estimate, we use a temperature stress factor adapted from the Arrhenius equation-which includes this term:
eEa/k(1/Tu – 1/Ts)
Ea = Semiconductor activation energy
k = Boltzmann’s Constant
Tu = use temperature (K), or the die temperature in the design.
Ts = stress temperature (K) used in the accelerated life test.
We control the stress temperature and it is based on the semiconductor process, generally either 150 degrees C (423K) or 175 degrees C (448K). You control the use temperature. Lower operating temperatures result in higher reliability. That’s your part of the job.
So, let’s say you want to increase the reliability of a system. Using the free MTBF tool referenced at the end of this blog, we can see the predicted effect of reducing the operating temperature from 100 degrees C to 90 degrees C.
At 100C, the calculated FIT is 1009.
At 90C, the calculated FIT is 860.
Is that enough of an improvment? That depends on your needs.
Please note I am not saying anything about the fundamental reality of these numbers. They are numbers, but there is a fair amount of speculation and unconsidered factors in a real life design.
It’s getting late and I notice there is about a half a glass left in the bottle. I suppose I could put the cork back in and save that last bit for later.
The unpublished papers referenced below are available by request.
The author would like to thank Thomas Welch and Raymond Oakley who collectively contributed three unhelpful suggestions and two rude comments during the writing of this article.
On Acceleration Factors used in Failure Rate Prediction – Unpublished paper by Thomas Welch, Director, Quality and Reliability, Fairchild Semiconductor
Secrets of Mean Time Between Failure – Unpublished paper by Raymond Oakley, Staff Customer Quality Engineer, Fairchild Semiconductor
Free MTBF tool from Advanced Logistics Developments, Free MTBF Tool
Failure Mechanisms and Models for Semiconductor Devices, JEDEC Publication, JEP122E
Reliability by Design, A. C. Brombacher, John Wiley and Sons
Reliability, Maintainability and Availability Assessment, Mitchell O. Locks, Hayden Book Company
Collective Electrodynamics, Carver A. Mead, The MIT Press
The Novikov Conjecture, Matthias Kreck and Wolfgang Lück, Birkhauser Verlag
This post is also available in: Chinese (Simplified)