First, National Semiconductor buys Fairchild Semiconductor, then spits then out before they are acquired by TI, now On Semiconductor wants to work them over. These mergers and acquisitions tell me that the epoch of non-linear growth is over in the semiconductor industry.
First, cell phones ate EVERYTHING. So, the only volume is in cell phones. Cell phones only need a few different types of chips but they need them in massive quantities.
Second, the cost of creating a chip is too high. The "maker" culture relies on the fact that creating a PCB is cheap. Between the cost for EDA VLSI tools and the NRE for creating a mask set, you blow $100K before you even get started making a chip. I suspect that we probably have the ability to digitally image 1.0um or larger features so we don't have to cut a mask set, but it's a chicken/egg problem. Since nobody wants 1.0um+ processes anymore, nobody will develop the technology for them, so nobody will design for them, so nobody will want 1.0um+ processes.
Third, modern fab capacity is ridiculous. A single 300mm wafer can produce 70,000 1mm^2 chips or 700 1cm^2 (a big chip nowadays). How many customers can consume those kinds of volumes?
> First, cell phones ate EVERYTHING. So, the only volume is in cell phones.
Not sure where your getting this. The MCU market dwarfs the cellular market, last time I looked it was on the order of 25B MCUs sold per year and the majority are still 8bit MCU. There are only ~1B cell phones produced per year. I don't think the discrete market is really all that driven by cell phones.
The problem with you argument is that 8-bit MCU's are a rounding error.
The Cortex A series cores are around 2mmx2mm. The Cortex M series are something like .2mmx.2mm. It looks like my 10mmx10mm comment was way off.
So, let's assume that the 8-bit MCU's are .1mmx.1mm. That's probably WAY too big, but let's start somewhere. That means that you can fit 7,000,000 on a wafer. For roughly 3,500 wafers for the entire, worldwide capacity of embedded MCU's.
The entire, worldwide, annual market for MCU's is 3,500 silicon wafers per year. That's the same number of wafer starts that a big fab has IN A WEEK.
Modern fabs can pump out a million a week:
electroiq.com/blog/2014/01/tsmc-samsung-and-micron-top-list-of-ic-industry-capacity-leaders/
This is why everybody wants to produce bigger chips like SOC's and why Intel will never switch to ARM until it's about to go under. If your chip is too small, you can't move enough to fill your fab even if you give them away free.
What will be the impact of this on jobs? I see a lot of layoffs in the semiconductor industry. I don't know what is the impact on jobs for a firmware engineer. I don't see anything new coming soon in the embedded space and IoT looks like more of a hype than anything else.
You have to distinguish between the wearables/smart home stuff and industrial IoT that has a real impact on cost saving and efficiency. Much of the former is idiotic. The latter is a huge market even if a lot of the details are still being worked out.
> I don't know what is the impact on jobs for a firmware engineer.
Moore's Law slowing down is good for firmware people. Optimizing, by hand, for 10% matters when you can't just get a factor of 2 automatically every 18 months.
> I don't see anything new coming soon in the embedded space and IoT looks like more of a hype than anything else.
The embedded space is power and cost constrained. Power constrained because of batteries and their inability to deal with RF. Cost constrained because memory now makes up 75% plus of chip area which defines the cost.
IoT is a lot of hype. The problem is that IoT is LOW VOLUME, and nobody wants to deal with that. So, everybody wants you to adopt their "highly scalable" (har har) backend and extract rent from the people doing the real work who are creating and marketing the application device. And, everybody decent at producing devices is also smart enough to realize that the money is in the backend so they refuse to play along.
The company that is willing to slog through the swamp of low volume will be the one who stumbles into the high volume application. Think Qualcomm getting started by tracking trucks and stumbling into cellular phones.
I'm not really sure how this pertains to On/Fairchild. On produces a few MCUs but as far as I know it's not a significant part of their business.
My point was that there are massively more applications that require supporting electronics, which is what these two companies specialize in, than there are cell phones built.
I also don't think your size numbers are realistic. You're not building Cortex A57's on the same process as Microchip is building PIC18s. Completely different size scales and completely different fabrication lines.
It's not like Fairchild/On could just 'pick up' a contract to build SoCs on the lines they use to build PWM motor controllers.
It's an interesting combination. ON would become the predominant manufacturer of discretes. There's still competition at IR, Vishay and Infineon (and ST if you can get past their horrible datasheets). But it doesn't really add to their portfolio, just takes away a competitor.
I haven't done the math but it doesn't seem like a great deal or a good plan.
16 comments
[ 2.6 ms ] story [ 58.8 ms ] threadUsually it doesn't work out too well.
First, cell phones ate EVERYTHING. So, the only volume is in cell phones. Cell phones only need a few different types of chips but they need them in massive quantities.
Second, the cost of creating a chip is too high. The "maker" culture relies on the fact that creating a PCB is cheap. Between the cost for EDA VLSI tools and the NRE for creating a mask set, you blow $100K before you even get started making a chip. I suspect that we probably have the ability to digitally image 1.0um or larger features so we don't have to cut a mask set, but it's a chicken/egg problem. Since nobody wants 1.0um+ processes anymore, nobody will develop the technology for them, so nobody will design for them, so nobody will want 1.0um+ processes.
Third, modern fab capacity is ridiculous. A single 300mm wafer can produce 70,000 1mm^2 chips or 700 1cm^2 (a big chip nowadays). How many customers can consume those kinds of volumes?
Not sure where your getting this. The MCU market dwarfs the cellular market, last time I looked it was on the order of 25B MCUs sold per year and the majority are still 8bit MCU. There are only ~1B cell phones produced per year. I don't think the discrete market is really all that driven by cell phones.
The Cortex A series cores are around 2mmx2mm. The Cortex M series are something like .2mmx.2mm. It looks like my 10mmx10mm comment was way off.
So, let's assume that the 8-bit MCU's are .1mmx.1mm. That's probably WAY too big, but let's start somewhere. That means that you can fit 7,000,000 on a wafer. For roughly 3,500 wafers for the entire, worldwide capacity of embedded MCU's.
The entire, worldwide, annual market for MCU's is 3,500 silicon wafers per year. That's the same number of wafer starts that a big fab has IN A WEEK.
And that was 2008. http://www.eetimes.com/document.asp?doc_id=1169357
Modern fabs can pump out a million a week: electroiq.com/blog/2014/01/tsmc-samsung-and-micron-top-list-of-ic-industry-capacity-leaders/
This is why everybody wants to produce bigger chips like SOC's and why Intel will never switch to ARM until it's about to go under. If your chip is too small, you can't move enough to fill your fab even if you give them away free.
Not by too much. A mobile baseband SoC that supports 2G+3G+LTE-Cat-6 will end up in the 7mmx7mm range with a 28nm process.
You have to distinguish between the wearables/smart home stuff and industrial IoT that has a real impact on cost saving and efficiency. Much of the former is idiotic. The latter is a huge market even if a lot of the details are still being worked out.
Moore's Law slowing down is good for firmware people. Optimizing, by hand, for 10% matters when you can't just get a factor of 2 automatically every 18 months.
> I don't see anything new coming soon in the embedded space and IoT looks like more of a hype than anything else.
The embedded space is power and cost constrained. Power constrained because of batteries and their inability to deal with RF. Cost constrained because memory now makes up 75% plus of chip area which defines the cost.
IoT is a lot of hype. The problem is that IoT is LOW VOLUME, and nobody wants to deal with that. So, everybody wants you to adopt their "highly scalable" (har har) backend and extract rent from the people doing the real work who are creating and marketing the application device. And, everybody decent at producing devices is also smart enough to realize that the money is in the backend so they refuse to play along.
The company that is willing to slog through the swamp of low volume will be the one who stumbles into the high volume application. Think Qualcomm getting started by tracking trucks and stumbling into cellular phones.
My point was that there are massively more applications that require supporting electronics, which is what these two companies specialize in, than there are cell phones built.
I also don't think your size numbers are realistic. You're not building Cortex A57's on the same process as Microchip is building PIC18s. Completely different size scales and completely different fabrication lines.
It's not like Fairchild/On could just 'pick up' a contract to build SoCs on the lines they use to build PWM motor controllers.
I haven't done the math but it doesn't seem like a great deal or a good plan.