Why, How and What of Custom SoCs
Ever watched the Ted talk by Simon Sinek How great leaders inspire actions? Not yet, then I encourage you watch this 20 mins talk. This video covers the most fundamental thing that most companies fail to address: connecting with customers! Often companies focus on their products, going into details about the technical features, price, engineering innovation, etc. However, they fail to address the basic thing that is needed for a successful sale: Why they are offering the product? Answering this question bridges the gap between product and market. Revenue is an outcome, not the sole purpose of a company’s existence.
Let us take an example of a conventional sales pitch for the embedded computing platform: System on Module (SoM).
“We offer SoM that has a SoC, memory, power circuitry, Operating System, and BSPs, all integrated on a small form-factor board that offers you a platform for building your next embedded product”.
Sounds exciting? Well, it depends. However, it does not generate a great interest. Now, how about the following as a sales pitch?
“Would you like to accelerate time-to-market for your next embedded product development, while also reducing development cost and risk? Well then, we have something that may interest you”.
The latter pitch nails it, by generating curiosity and interest, yet the funny thing is that we are yet to use the term ‘SoM’, which we are supposed to sell! It starts from the customer-side, instead starting with tech jargons.
Custom SoCs are not new in the semiconductor market. However, with latest industry dynamics and enablers, there is a renewed interest in custom SoCs. As someone who is not an expert either in technology or in marketing, but maybe a jack in both 😊, this post is my humble attempt to apply Simon Sinek’s approach to custom SoCs. Hope you enjoy this post and share improvement areas on the same.
Let us explore some of the industry dynamics, customer expectations, and motivations, for the push towards custom SoCs.
Questionable Moore’s Law
For more than five decades, Moore’s Law has guaranteed performance boosts, yet at lower cost and power consumption. With time, performance enhancement and power optimization was obvious. However, currently the economic equation, which is guaranteed by this law, is failing. With leading process nodes tending to reach atomic level, the designs are becoming complex, leading to long time for commercialization, thus the cost equation does not hold true. With custom SoCs, performance boosts and power efficiency is possible without compromising on cost.
Frugal and cost-sensitive IoT end-nodes
With smartphones adoption flattening out, IoT is the next wave that will drive revenue and growth of the semiconductor industry. The problem is that IoT is not a product; it is an ecosystem of products and services, each playing a specific role in the value chain. The use-cases are not dictated by the supply-side, but by the end-users’ (or the demand-side) need for optimizing cost and increasing productivity, efficiency, safety & convenience. For e.g, an industrial plant intends to reduce down-time of the system. They can put IoT end-nodes at various points, and whenever there is a deviation from the expected behavior, the relevant end-node sends alert, so the breakdown gets repaired in least possible time. The data collected from all the end-nodes can also be used to perform analytics, which helps to explore means to predict failures and increase efficiency & productivity.
Most of these end-nodes will be customer-specific and application-specific, and work in ambient conditions depending on the type of industry. Stock SoCs can be an over-kill for these niche applications, and may not fit within the budget of many start-ups and small companies working on specific IoT use-cases.
Stock SoCs are mostly a commodity as any company can use those and build their products. So, competitive advantage and differentiation are limited to mostly software. However, with custom SoCs, companies can extend the differentiation into low-level hardware as well. Some proprietary IPs, accelerators, etc. can be integrated on the SoCs to enhance performance and power efficiency.
Stock SoCs are like buffet meals, in which you pay for the entire course irrespective of your appetite and food preference. As true for ala-carte option, in which you can reduce meal cost and order as per your choice, with custom SoCs, BoM cost can be reduced as you will integrate only parts that are needed for your end-product. Some reports mention that BoM cost and die size can be reduced substantially with custom SoCs.
Artificial Intelligence (AI) and Deep Learning (DL)
AI and its subset DL, Machine Learning (ML) has the potential to offer enormous possibilities. Few applications can be NLP, Computer vision, automation, etc. Currently, GPUs are mostly used to address the computation needs for AI, as the former can perform parallel processing on humongous amount of data, needed for training and inference. In future, we will see many applications and products that use AI for enhancing performance and efficiency. Many companies are venturing into custom chips that are optimized for AI. Tesla is supposed to be working on custom chips, which may have state-of-art computer vision and machine intelligence, for realizing driverless cars. TPUs from Google are optimized for AI, and some reports mention that the 2nd gen TPUs can outperform Nvidia GPUs for AI tasks. Apple moved away from stock GPUs, and planning to build GPUs in-house, maybe to integrate more AI tasks on future iPhones.
In this section, I will focus on the enablers of custom SoCs. There are many proprietary ISAs that offers architectural license to build custom SoCs. However, I will cover only the two factors that are enabling penetration and adoption of custom SoCs, at low cost.
RISC-V ISA deserves a special attention for giving a strong boost to custom SoC market. Being an open-source ISA, RISC-V will enable designers to build custom SoCs at very low upfront cost. Designers don’t have to pay for the license and royalty. This is sort breakthrough innovation in the semiconductor industry, as it will add many new customers for custom SoCs. RISC-V is gradually building an ecosystem with IPs, debug-tools, etc.
The ARM DesignStart enables designers to build custom SoCs at low cost. It offers free of cost access to Cortex-M0 IP, EDA tools and physical IPs. With a minimal cost, designers can license the ARM IP for production devices.
Finally, I would like to touch on the technical part. With my limited understanding, my view is that custom SoCs are disrupting the existing fabless semiconductor value chain. Usually, OEMs procure stock SoCs offered by the SoC design companies such as Qualcomm, Nvidia, NXP. The SoCs are designed based on some licensable CPU IPs (ARM, MIPS, etc.), then manufactured at fabs such as TSMC, Global Foundries. SoC design companies attempt to address multiple markets and applications with their stock SoCs, as this strategy offers benefit of economies of scale. Custom SoC is breaking this value chain, with OEMs bypassing the SoC design companies, and dealing directly with IP vendors, EDA vendors and fabs.
With custom SoCs, designers can integrate any proprietary IPs, analog sensors, mixed-signal parts, DSPs, accelerators, etc., depending on their application needs. This also enables designers to push many tasks to hardware, instead of doing the same in software. The benefits of pushing things to hardware are more security, reduced software memory footprint, better performance and low power consumption.
I strongly believe that emerging market of IoT and AI will provide a huge thrust to custom SoCs. In the last decade, stock SoCs has seen unprecedented growth with smartphones. However, in future we will see more IoT and AI applications integrated on smartphones and other embedded devices, and then maybe custom SoCs will steal the show from stock SoCs. General-purpose/ Stock SoCs will still play a dominant role; however, we will see more industry focus on application-specific parts.
It will be something like going back to past again. The industry started with ASIC, and then moved towards stock SoCs for cost advantage, and now we are seeing some motivation towards going back to custom SoCs (sort of ASIC) motivated by cost, performance and power efficiency.