Smartphones play a vital role in our life; it is sort of panacea for us. It has achieved massive growth in the last decade. With such demand of the end-product, it is obvious that the supply side has worked in overdrive. Semiconductor industry has seen massive growth due to smartphones. The number of OEMs increased to fulfil the volume demands, pulling more vendors into semiconductor industry to suffice the OEMs’ supply chain requirements.
Few OEMs made fortune in smartphone sales, with revenue exceeding billions cumulatively. Such capital inflow encouraged them to invest more money on product development, to fulfil the end-users’ paradoxical requirement of more performance with more battery life (less power consumption) at a lower price. With cumulative silicon sales going into billions, the semiconductor industry responded positively, and focused on stretching the innovation into leading process nodes and other techniques to enhance performance.
Smartphones as a forerunner
Smartphone fuelled innovation in the semiconductor industry. System on Chip (SoC) vendors were excited about the huge sales prospect and did their best to capture the market, by venturing into expensive leading process nodes, complemented with research to enhance performance without compromising on power consumption. Obviously, smartphones acted as a forerunner in terms of technology adoption. With smartphones’ shelf life of anywhere between 18-24 months, SoC vendors churn out new offerings with less lead time, as each SKU has guaranteed sales volume of millions.
General-purpose or stock SoCs achieved humongous growth in terms of sales and innovation. All these innovations have some externality, as these latest technologies spilled over to adjacent markets such as embedded and consumer electronics, and led to their growth.
Is the party over?
Currently, the smartphone growth rate is slowing down, thus the silicon demands from the OEMs will also reduce. Commoditization has also kicked in the smartphone market, so most OEMs are struggling, as end-users focus on price instead of brand.
With such industry dynamics, many chip vendors, who have just relied on smartphone as their majority revenue source, will face tough times ahead. We have already seen Imagination going off the block, after Apple pulled out of using Power VR GPUs for upcoming iPhones, maybe Apple is taking its differentiation strategy to next level. Read my related post on this: Apple, Imagination and beyond …
Has smartphones lost the slot of being the harbinger of technology?
The road ahead
With my limited understanding, smartphones as a market has lost its sheen. As the demand from the end-users will plummet, the semiconductor industry will respond accordingly. I believe that smartphones has lost the technology forerunner place to two new markets namely: IoT and AI.
Currently, the semiconductor industry is excited by the avenues of both these markets. Silicon companies are trying to associate with these technologies in some way. You can check the marketing collaterals or website of any big or small company in the chip industry, most of them will have some mention of IoT or AI, or maybe a combination of both. We are seeing a transit from a product-centric era (smartphones) to an application-centric era that encompasses the extensive usecases and applications feasible with IoT and AI. Both these technologies are placed at extreme ends on the performance scale.
IoT is not a product; it is an ecosystem of product and complementary services working in tandem to bring efficiency and optimization in any industry. Silicon sales will be largely driven by billions of low-performance IoT end-nodes or sensors that will be spread ubiquitously to collect ambient data, and then pass it on up the value chain for analysis. These end-nodes are frugal cost-sensitive devices with low power consumption.
At the other extreme of performance scale are the AI applications. These involve analysis and inference of humongous amount of data. AI can find use in diverse domains including driverless cars, ADAS, NLP, Computer Vision, Image recognition and many more. These applications need analysis of humongous amount of data for training and inference. So currently, GPU, with its parallel processing capability, are used extensively for these applications. Architectural innovation such as NVidia’s CUDA enables outsourcing of AI tasks to GPU, while the CPU can carry out system related tasks. This optimal combination of GPU and CPU on a SoC, with CUDA integrated, can be of great virtue in future, as we see AI being integrated into many applications. I believe that the industry is also getting prepared for AI. Most of the leading companies including Google, Apple, Microsoft, Intel, Facebook, Baidu, Qualcomm, along with many startups (THINCI, Groq, Nervana, etc.) are exploring the ideal way further on building AI capabilities. Hot Chips 2017 was dominated by AI and ML topics. Check a piece on EETimes here.
Edge computing, which involves processing data near to the source, is also gaining momentum with AI-powered chips, as opposed to a conventional approach that pushes data to central servers over networks, leading to transmission delays. Edge computing is useful for applications, in which real-time decisions are needed, such as driverless cars, ADAS, robotics arms, etc. AI integrated in smartphones will be one more usecase of edge computing, as this will offer real-time user experience.
Smartphones integrated with AI capabilities are work in progress. Apple is working on it. The next smartphone from Huawei will feature in-house Kirin 970 SoC that has AI processing capabilities. I am sure soon other smartphone vendors will be launching products with AI features integrated. On the other side, the power optimization innovations, which are happening on the IoT end-nodes, may be used in smartphones to extend battery life. So, smartphones, which was earlier driving technological advances in other markets, is now integrating technologies from other markets and applications.
The next technology frontiers will be driven by AI and IoT market. We will see chip industry responding to the needs of these markets by offering products that are specific to these markets. The IoT market will look for cost-sensitive and power-efficient platforms; the AI applications will be inclined for high-performance, yet power-efficient platforms. General-purpose or stock SoCs may not be ideal for meeting the requirements of IoT and AI applications. So, the industry will look beyond stock SoCs.
IoT and AI are not mutually exclusive. In certain applications, both IoT and AI can be used together. The IoT end-nodes will collect ambient data and passively pass it up to the server, which has AI capabilities, to analyse and understand the data, and then take some actions.
The inflection point
Stock SoCs have seen massive momentum due to smartphones in the last decade. However, smartphones’ sales are flattening out gradually, thus the sales volume of stock SoCs will fall southward. Simultaneously, the AI and IoT markets are picking up, and these markets need application-specific chip, not a general-purpose one. So, along with stock SoCs sales fall, we will see positive interest for custom SoCs or ASICs.
Demand for custom SoCs has also opened up avenues for small companies and startups to enter into chip design. The entry-cost barrier for custom chip design is reduced by programs such as SiFive DesignShare and ARM DesignStart. ARM DesignStart is ideal for building custom SoCs for IoT market, as the program includes low-performance Cortex M0 and M3 IPs. I didn’t see any specific markets focus for SiFive DesignShare, so maybe this program is application-agnostic. With less capital expense, custom chip design can be done. The only barrier is the high fab cost associated with production. For IoT, it is not much of a concern, as matured nodes can be used. However, AI chips will need leading process nodes for maximum performance at least power consumption. To know more about custom SoCs, take a look at my earlier post Why, How and What of Custom SoCs.
RISC-V is democratizing custom chip design with open-source ISA. SiFive, the pioneer of RISC-V, offers DesignShare program that fast tracks custom chip development, at low cost. The program relies on partners, which includes semiconductor and custom-silicon companies, to offer low-cost IPs for designing custom chips. Check my related blog – ARM vs RISC-V: A Game Theory perspective .
I do not assert that stock SoCs will lose market share completely. There are many other markets for stock SoCs. However, the sales volume of stock SoCs will definitely fall, as smartphone adoption slows down. Other markets cannot compensate for the shortfall in silicon sales.
How this is going to impact the SoC vendors, whose dominant end-market is smartphones?
Chip industry is capital-intensive. With large sale volume, SoC vendors can leverage economies of scale to optimize the Cost of Goods Sold (COGS). However, as the smartphone sales flatten and the industry gravitates towards application-specific products that are based on custom chips, vendors have to look for market beyond smartphones. AI and IoT are the next emerging markets. In my view, maybe not all SoC vendors can diversify their chip portfolio to cover these markets. Falling revenue will encourage the pursuit of diversification, so I believe we will see more M&A in the semiconductor industry. We are already seeing Qualcomm pursuing NXP, which is a dominant player in IoT.
Smartphone OEMs, who have backward integrated into in-house chip design, are trying to differentiate their smartphones with innovation at the chip level. Apple and Huawei have plans to launch smartphones with SoCs that supports AI. I believe Samsung and Xiaomi will also join the party soon.
Summing up, the chip indutry is in a state of flux now. As Moore’s Law is challenged by both technology and cost factors, performance enhancements are not anymore an obvious output of time and investment. We will see some real innovations happening that goes beyond the chip, and focus on the system as a whole.
I believe the post relies on many assumptions that are based on my limited understaning, so flaws are obvious. I look forward for any flaws and improvement areas.