Does Integration improve Reliability?
The topic of integration vs. reliability has always interested me, and now more so since I am invited to be a panelist on “Does a high level of integration make power converters more reliable or not?” at the upcoming industry power electronics conference (APEC). There are learnings from computer history, and I got to thinking how we got from ENIAC, which had a failure rate of twice a day in 1945, to today’s modern microprocessor with billions of transistors shipping in volume with high reliability.
Making components more reliable certainly helps. The ENIAC engineers were able to reduce the failure rate by 4x with more reliable vacuum tubes, and the next level of reliability came with the use of the transistor. Early transistor computers were built from Ge, for example, TRADIC in 1954, which was reliable enough to be installed in a B52 airplane. Ge transistors, however, did not handle heat well, and the next step was achieved by Gordon Teal at TI in 1954, by figuring out how to make transistors out of silicon.
Reliable transistors alone do not lead to reliable computers, since they also need to be interconnected. As computers grew in complexity, engineers ran into the problem of tyranny of numbers which can create all kinds of reliability issues due to interconnecting the many components. The invention of the IC in the late 1950’s by Jack Kilby at TI and Robert Noyce at Fairchild solved this problem by allowing one part to be made with numerous interconnected components. Over the years, improvements in manufacturing and lithography have taken the transistor count from 2250 in the Intel 4004 microprocessor in 1971, to 16 billion in the Apple M1 presently. Integration has made microprocessors reliable because a single part is manufactured out of a large number of interconnected components using a scalable, precise and well controlled process.
Power transistors are made by interconnecting a large array of smaller transistors and also considered a single part. While it is to be debated whether a high level of integration makes power converters more reliable, it is clear that modern power transistors are highly integrated and reliable.
Nice article, Sandeep. Yes, the key is the integrated circuit. The caption of your illustration says that the Cray1 supercomputer contained 50 miles of wiring. I have to admire the Cray engineers of 1975 for such a feat! But let's compare what we can achieve today. The standard field size of a lithography tool is 26mm x 33mm. With EUV lithography, we can print (still in R&D but hopefully soon in HVM) dense lines with a pitch below 30nm. OK, let's relax a bit and print dense lines with a pitch of 33 nm, and your (big) chip takes up the entire exposure field and you fill the entire area with a long wire in a serpentine fashion. This is the wire's total length: 26mm x (33mm/33nm) = 26 kilometers. And this is only one layer in a chip. So, Long Live Moore's Law!