TI currently supports about 78 different process technologies, more than 20 of them inherited with its 2011 acquisition of National Semiconductor. "It's a whole different world in logistical complexity," he said.
Wow. I for one find it kind of unbelievable that TI can support so many different process technologies. And I think Menon's quote, about being a whole different world in logistical complexity, has got to be an understatement. You would thing they would want to cut that number down just for complexity's sake.
"Wow. I for one find it kind of unbelievable that TI can support so many different process technologies."
I would say that that depends upon how different each process technology is. Is each of the 78 a totally seperate process, or an interation of an exisiting process? You figure they have 6 or 7 geometry nodes and then 10 different tweaks to a process on each node and probably some overlap between what TI and National processes that are more or less the same?
That's a good point. Obviously each of the 78 are not completely different processes. There would be considerable overlap, I imagine. And of course, as you say, TI is working across a lot of different process geometries requiring different processes. Even so, 78 sounds like a lot.
It does sound like a lot, too many to manage effectively really. But suppose I tell you I have 78 cents, but what I really have is two quarters, two dimes, a nickel, and three pennies. Technically I have 78 cents, but I only have 8 coins to carry around - much more managable than 78 which would likely put a hole in my pocket ;<)
To put some perspective on the "78": TI (and the other giants in analog) has (directly or indirectly) absorbed many other analog IC companies, and continues to manufacture/support a lot of those products I am sure. Think about Unitrode, Burr-Brown, Philbrick, etc., each with their own markets and processes optimized for them. It's undoubtedly just simpler and more economic for TI to keep those fabs up instead of trying to force the designs into a different process! That's just another way of explaining the "tweaks."
@Rick, Can such information be gathered about other Analog giants like ADI and Linear tech ? I suspect those companies will have as many or more proccess technologies. It will also be interesting to find out how many of these companies still engage in development of SiGe Bipolar/Bipolar process technologies crucial for RF and Microwave circuits.
I agree with the statement - there is no Moore's law for analog. Smaller geometries do not necessarily benefit analog or mixed-signal devices. The issues created by shrinking an analog device often outweigh the benefits.
@Rick: it is interesting that the statement "Some processes aim to support 600-700 volts, some aim to support a wide array of voltage breakdowns..." steers clear of mentioning anything about non-Silicon semiconductor processes in TI's technology. It would have been nice to know if TI is working on SiC, GaN tecnologies, both in native substrates and in combination with Si substrates.
Based on experience at Motorola over many years, I would agree that 78 flows can be managed, but across several fabs, and analog priorities are precision, power, and voltage rather than density since Moores Law only applies to low power and voltage, and often must make some compromises for precision. CMOS was not around during early days of analog-mostly IC development but gave the standard memory and processor guys a way to "stay on Moores Law" plan.
However, the real world is analog and so SOC chips have to mix digital and analog, and the sad news is that SOC's will not follow Moores law for that reason (cost wise) but they can appear to be on track for their "smallest critical features." Marketing, not engineering. We even tried to put high voltage verticle devices and microcontrollers on same chip, and found that best economic solution was a 2 chip module. But there is room for MODERATE analog power and voltage mixed with digital memory and logic, DMOS-CMOS.
So let's hear from the SOC gurus on cost-benefit learning curves. Just give it another name. And if time permits, see if you can find any mixed-signal stacks of chips in volume production yet.