The assumption the author has made is that each chip has an R&D spend proportional to the revenue that it generates (~20%). This is clearly false. Some chips are relatively simple but sell in huge numbers so the R&D per revenue is extremely low. An SDR is one of the most complex chips you can make with a wide range of expertise going into the design. I can easily imagine $20+ million being spent on R&D for this chip and the market for it is pretty limited (not for consumer products). Even at 100,000 sales the R&D costs alone would be $200 per chip.
Pricing for these sort of niche low volume products is guided much more by what the market tolerates paying rather than the production and development costs. A large number of these types of chip never recover their R&D spends.
>> I can easily imagine $20+ million being spent on R&D for this chip
I would personally estimate around 30 - 50% of that figure. Whatever the number is, it's really funny money in a company such as ADI which sells a large chip catalog. Most of the modules in that chip can be reused in others, and many of them would themselves be reuses of modules from other chips, perhaps with some modifications. So the total R&D cost would in fact be spread amongst many projects and it would be pretty difficult to nail down exactly how much is any one project's share.
To add to this, I imagine chip companies don't look at it as plainly as whether a chip recovers its development costs. If the development costs lead to techniques or designs that can be reused, the next chips will be not only better (due to incremental improvements) but also cheaper.
In my experience designing electronics, EEs and the people handling logistics/manufacturing are extremely risk averse actors who make many decisions based on trust and personal relationships. As a contractor, my deadlines would often only allow for at most two or three prototypes before I had to deliver a working design and that did not leave much room for error. If my client or I had an existing relationship with a vendor that always supplied parts on time, fast samples/dev kits, or if I knew their reference designs to be more reliable than most, then that vendor would be the preferred source for parts even if their other offerings weren't the best or the most cost efficient.
You can also often get great pricing for low/medium volume orders if you have built up a history so I've had situations where I had to design a significantly more complex board with a microprocessor that normally cost 10-20 times more than the CPU I would have chosen (from another vendor). The client's supplier for the other most expensive part on the board had too much inventory of a soon to be discontinued processor (that was pin compatible with a newer one, also very expensive) and the account manager's boss was willing to sell us everything we needed at an 80% discount off high volume pricing with pricing guarantees on the newer parts for several years, all to meet some regional sales quota.
As another example, microprocessors often have arduous power up requirements where different voltages have to be supplied in a specific sequence within annoyingly specific timing tolerances. In a project using Marvell microprocessors, it was much faster to just use Marvell's power chips and existing reference design than to waste precious time on modifying designs for cheaper chips from other vendors. The power parts cost five times as much as they would had I chosen my favorite vendor but because the board was a mixed signal design it was extremely expensive to make changes and test prototypes.
A vast catalog of parts is an especially powerful market force if you are a company like TI that has a well developed software tool like WEBENCH which can spit out dynamic reference designs for most of their power chips. Unless I have an existing reference design from a vendor of a major part (like the CPU above), I usually go straight to TI's tool to design any power supplies I need so I can focus on the meat of the PCB. If TI also has parts that do something else I need on the board I will often just choose them instead of researching other vendors (unless it's a really expensive component) and that compounds as I make more and more PCBs, each time becoming more comfortable with TI parts at the expense of others.
You can calculate a sort of "time-independent cost" of R&D by taking any "innovative elements" that were discovered during the project, and retrospectively amortizing their costs across all the future projects one expects to use them in (maybe with NPV applied to the weightings of the amortization to represent lost opportunity-cost of having that money sit in an investment.)
This is not a niche, low volume IC. Its smaller brother is used (I should say was used) by DJI in its Lightbridge video transmission system (that you could find in the Phantom 3/4, that sold in the order of millions). And that's only one product I know of, I'm sure there are many other.
I would love to know how much DJI was paying for that chip, especially after they went to the effort or replacing it with an in-house one (no name Chinese startup) in more recent batches.
They have been surgically copying/reimplementing advanced technology for over 10 years now. While its still mainly for internal consumption (cars, electronics), western suppliers are steadily pushed out of more and more niches, and some products do get exported (LCD panels for example).
Is it accurate to refer to the average wafer price as "manufacturing cost" rather than "materials cost"? Shouldn't the amortized cost of the fab be included as part of "manufacturing cost"? Or is that captured in any part by "R&D"?
Either way, the drug comparison is apt. The chip/pill is a tiny thing with negligible marginal cost. The plant costs billions. Pricing seems arbitrary, but these huge expenses do need to be recouped somehow.
To eliminate fab cost from the equation you could price out what it would cost to get it built by a 3rd-party fab. It certainly wouldn't be as cheap as shown here, but still cheap enough to make the point of the article.
No, NRE is masking and tooling cost. Fab costs would need to be amortized separately. (Because fab costs get spread across many, many products; the 1000 wafers used as a guesstimate in this article is tiny compared to what a fab produces over its life.)
If anything, the Fab costs should be included in the cost-per-wafer.
Analogue fabs are highly customized. You could probably make some back of the napkin comparisons but you wouldn't be able to actually manufacture it in another fab without a lot more R&D spend.
The difference is with analogue everything is a full custom, with digital it's only the trailblazers. Also linear devices require an order of magnitude more control and characterization than digital even when full custom.
I think we're mostly in violent agreement. There's a lot for an analog IC - those mid-manufacturing test pads are a great example of pain. Testing costs are going to be higher, there will be extra NRE on custom test harnesses, and all the rest you allude to. But they'll be able to get multiple projects out of the basic investment in litho gear. I think it depends where one draws the line at the (very large) CapEx of "the fab".
Fab cost is the NRE because that's how a hardware developer is paying the fab costs. Most fabs' cost are spread across its many customers and products. They build it into whatever they charge for the runs. If AD owns the fab, then it's likely a separate organization in the company doing something similar. So, for purposes of AD9361 designers, the NRE includes the fab costs since they probably didn't build a fab as part of AD9361 project. Just paid one the NRE.
Not when you're estimating the cost to develop a specific chip. I can avoid the per-project NRE by simply not doing the project.
And, again, if a project produces more chips, then more of the fab production costs will be amortized across that project. The definition of Non-Recurring Engineering costs are that they are fixed and don't increase as the unit volume goes up: https://en.m.wikipedia.org/wiki/Non-recurring_engineering
The way the fab could be considered NRE is if the fab were built only for this chip. (Or if a fixed percent of it was). In advance.
It is worth mentioning that drugs are only so profitable because of their legal status. Drugs themselves are very cheap to make - additional cost is simply the risk of getting caught, security and so on. I think those factors don't exist for microchips, so the title is a bit misleading.
The title is very misleading. Had the author stuck to legit pill making, it could’ve made a fairly apt comparison, but they mentioned drug cartels which implies a very different set of costs / risks. I’d further argue drug cartels have a wider price differential between raw material and wholesale product too, but don’t have a source to back it up.
Turns out, if BusinessInsider is correct, the raw coca plant leaves needed to make 1kg of cocaine base costs about $650. The same article claims you can get kilos in South America for about $6,500 with purity around 85%. Prices go up the closer you get to the final destination, but as do the risks too of course. If those numbers are accurate, would indicate a gross difference of 10x, not the 172x from source article. I’m surprised the coca leaves go for so much!
Remember that final prices of drugs are “street prices” in quantities of use. So you might have 25,000 USD of a drug, and it might be 2 kilograms, but when you split it up into 2-4g packets, the individual prices are much higher and you might get “one millions” in street value, which is what the police and media like to report. It makes the amount seem so much bigger.
Yes, I know street value. I also know stateside wholesale price too. I just didn’t realize the markup was so huge between source and destination vs raw material and finished product.
Also, cocaine isn’t usually sold in 2-4g packets in most places. Usually it’s 1g or less. Otherwise (but much less often) it’s ~3.5g.
If you are an actual drug company, I suspect politics and approval processes are at least as expensive to the company, as the actual R&D itself.
It is incredibly hard to get into the market with todays legislation.
Drug cartels probably spent a lot of money getting out of legal trouble and dealing with international police
I'd argue that for drug development, clinical trials are part of R&D. After all, their point is to prove that the product works as intended and that it's safe. For any other product that's considered part of R&D.
Not all drugs are cheap to make. Biologics are notoriously complex to produce reliably and much more expensive than your typical small molecule drug encapsulated in a tablet.
Is this actually true though? I'm sure some now-generic drugs have pretty complex and hard synthesis problems.
I know some medications for sure have struggled because of manufacturing. Whether these are actual problems, or legal problems because some part of the the synthesis is patented, I'm not sure.
ADI is a remarkable company and they seem to be doing very well these days. If you’d told me ten years ago that in 2018 all of the old-line semiconductor companies (ADI, TI, etc) would be trading at all-time-high stock prices I would have thought you were nuts but here we are. Hardware is still quite valuable.
Too bad they destroyed the Hittite and LT web sites. Those were for engineers. The AD web site must be laid out by some marketing goob; totally painful to use.
At least the AD website usually works on weekends. Many of these companies’ websites apparently need constant manual correction to continue operating, with the result that i often find I can’t download datasheets on Sundays. Bad for the hobbyist for sure.
The Hittite web site had some great tools such as a graphical mixer spur calculator. Of course that is gone. ADI tech support sucks. The HMC6300 documentation is poor and buggy, and tech support won’t help. Used to be you could call Hittite and speak with an engineer. ADI is a black hole of web forms and forum posts.
I just hope they leave LT Spice alone, but I’m sure they will fix it until it’s broken.
What the process? The tear down says mostly FET based, so I assume a BiCMOS?
The AD9361 require front end filtering, as the 5th harmonic of the 2m ham band falls in a national LTE band; horrible interfence from that wide open front end.
It’s not the 5th being down. The 5th LO harmonic mixes the LTE band down to the 2m band. That’s just inherent in a Gilbert cell mixer. There is a high powered, 20 MHz wide (OFDM) signal pretty much everywhere at 715 MHz.
This chip has been a boon to SDRs. It moves the challenge of building an SDR to the filter banks and power amplifiers (the input and output conditioning circuits). Although these days most of the new transceivers have a low end in the 300Mhz range. It is somewhat easier to mix up the low frequencies into the 300+ Mhz range to cover them.
clickbait title. not even a good analysis. i do like that the author wasn’t in any way complaining about the price or arguing that it wasn’t reasonable. in electronics (compared to drugs) there isn’t unfair pressure in the marketplace. so double phoo on the title.
Since the frequency range doesn't match what the chip is capable of then probably not. The chip was released in 2013 so I'm not sure what you mean by timing.
Pricing for these sort of niche low volume products is guided much more by what the market tolerates paying rather than the production and development costs. A large number of these types of chip never recover their R&D spends.