A distinct and important trend is emerging and worth paying attention to in the semiconductor industry. As I noted a few weeks ago, a renaissance is happening in the chip industry. Semiconductor innovation is not just cool again, it’s extremely important and perhaps, more important than ever. For a small period of time, the need for more processing power was not abundantly clear in many use cases. During the height of the boom in the Intel era, it was clear how slow our PCs were as we watched software open in time measured by seconds instead of milliseconds. The smartphone era saw advancements in semiconductors that gave us faster cameras, better visual effects, faster internet speeds, and made needed advancements in silicon. Today, however, things like AI, visual computing, autonomous cars and robots, etc., show us the limits of the performance we have today in our processors and leave us wanting more. This reality has re-ignited a competitive race in silicon design again by the incumbents like Intel, Qualcomm, Nvidia, and AMD. But an important shift has happened from the last boom to this new one.
From General Purpose to Specialized Architectures
The best way I can describe this trend is to say we are moving from a general purpose processor design to one more custom and more specialized to its applications. Perhaps the best thing to read that articulates this trend is this post from Qualcomm, who is leading the pack on this shift in some important ways.
What Qualcomm explains is how they, a semiconductor design firm, are embracing this new reality where designing a general architecture for as wide of an application as possible no longer yields a competitive advantage. Fascinatingly, it is no longer competitive for semiconductor companies to design the best processor architecture to yield the fastest chipsets around. Performance has become commoditized and, in its place, the value has moved from the single part, the processor design, to the entire solution designed to compliment the processor. To put it simply, the value has moved from one part to the sum of its whole.
For this reason, all the major semiconductor firms are now designing architectures that are much more flexible. They may design a core CPU or GPU solution but include flexibility in the design to allow for some variation in what is printed on the SoC from things like a modem, camera sensor, memory, etc., as well as complimentary co-processors for things like audio, video, etc. The optimization of all these parts, designed to work together to access more specific applications and use cases, is where the focus of the best designers in the world is now spent.
Perhaps the most interesting part of this trend is how it has allowed non-traditional silicon design companies to start giving headaches to the traditional firms. The highlight example is Apple, who now designs so much of its own silicon that companies like Intel and Qualcomm keep losing more and more of their Apple business to Apple itself. This may not be a problem if Apple was not the only technology company who needs, and ships, high-end silicon in volume. In fact, Apple ships almost as much 64-bit silicon as Intel does in any given year and Apple uses it only for themselves where Intel sells to the entire PC and server market. Apple was increasingly becoming Intel and Qualcomm’s most prized customer monetarily about the time Apple began designing them out of their products.
Similarly, Samsung designs its own processors and is increasingly designing as much as they can. Huawei designs their own silicon for many of their smartphone and network products through their wholly owned subsidiary HiSilicon. Most recently, Xiaomi has begun making its own processors for use in its products. Every time one of these companies starts making their own silicon, it is a lost revenue opportunity in some way for Intel and Qualcomm (mostly Qualcomm in the examples I listed).
The paradox in all of this, especially in the case of Qualcomm, is how the very thing which allowed them to rise to dominance is the very thing creating the many headaches. I’m talking about ARM itself. Qualcomm has benefitted from ARM IP and has been innovating on top of that IP for a decade or more. Now that ARM is making it easy for more and more companies to start using the ARM IP for their own differentiated advantage, it is opening the door for this competition and, more importantly, this deep customization we are seeing from the likes of Apple, Samsung, Huawei, Xiaomi, etc.
ARM’s continual pursuit in breaking down barriers for companies to license their IP, in essence chasing down more licensees as new customers, is the underlying foundation of this entire trend and I expect it to continue. More and more companies will start designing either parts of their own chipset solutions or the entire thing. This is one reason why I’m watching ARM in the server space so closely. If, and this is a big if at the moment, ARM servers become something that can take meaningful share of the market, then this same trend we are seeing in mobile may come to the server space. It would mean it would not only be possible but even likely Facebook, Amazon, Google, and even Microsoft may start designing more and more of their own server chipset solutions in order to differentiate. To be clear, I’m suggesting that what we see happening in mobile with the biggest smartphone companies in the world making their own chipsets, we could see the same thing happen in server with the dominant cloud services providers designing their own solutions.
All of this is possible because of the reality that the processor itself is no longer the key player. The value of becoming the total solution, wisely configured for optimum performance in its intended use case, has broken old paradigms and enabled a new era of architecture design. This trend is under-appreciated but also, potentially, fairly disruptive.
6 thoughts on “An Important Trend in Semiconductors”
Wouldn’t Qualcomm also offer a “everything on it” version as well, thus making everything else a custom subset? I ask because:
a) I’m curious if there are technical barriers to the “complete” version. This would yield the no compromises broadest functionality.
b) I don’t thing it benefits anyone, other than the sellers to over-customize, thus limit functionality. This yields prematurely disposable devices. That is wasteful.
I say this because I wonder how different this is from the x86/x86SX stupidity. Basically the X86SX had the math coprocessor disabled. Having the X86 built in coprocessor hurt no one, but gave greater latitude as jobs or users changed. The SX approach allowed Intel to sell chips with defects in the math circuitry, to the “you don’t really do any math” crowd, thus increasing yields. A user should be able “to do math” and not have to make that decision up front.
The question is What is the “everything on it” option? If they keep adding things to the processor then surely that will have a power drain issue.
When Apple builds their custom chip they build the features in that are needed whether that’s the fast processors for when they’re needed or a pair of low power chips that can be called upon when processing power is not needed or special DSP stuff to process images.
But the point is that everything that is required to drive the new phone is in place. On a single chip.
If Qualcomm have an “all you can eat” version of the processor the vast majority of customers – manufacturers – probably don’t need all those options.
Everything on it would be the one with all the features they offer. Yes, it may or may not drain power. Components could be off until needed, or a bigger thicker model could be offered, say iPhone 3GS type thickness.
I would prefer (for myself) to not have to choose which features I want, I would want all of them.
I’m not convinced. You appear to be speccing up something that the vast majority don’t need and if it is a small run then it will be expensive therefor not much use.
Nice in theory but lacking in practice.
There are many, many things most people don’t do. Gratefully other people do them. Yes, I like the most spec. I’m not speaking for most people, I’m speaking for me. Prefer does it but I don’t usually, to doesn’t do it at all.
“If they keep adding things to the processor then surely that will have a power drain issue.”
Power gating makes it possible to have parts of the chip that are not in use turned completely off. So power drain from more on chip features is a relatively minor issue. The real issue is that more features = more square millimeters of silicon = more cost.
A big versatile chip is going to be an expensive chip, and the market for devices with big expensive chips (ie, flagship phones) is pretty much limited to first world buyers. The market for small inexpensive chips, OTOH, is the entire rest of the world.