SoC + AI = SiPx

The market for third-party semiconductor intellectual property (SIP) continues to exhibit growth well beyond the (CAGR) Compound Annual Growth Rate for the semiconductor industry. Semico just completed an in-depth analysis and breakdown of the SIP market in a report called Licensing, Royalty and Service Revenues For 3rd Party SIP (SC105-18). The 2017 to 2022 CAGR is projected to be 10.9%, about 1.5 to 2.0 times the overall semiconductor market growth.

In order to have double-digit growth you must have drivers sufficient to propel the SIP market, and Semico believes the market certainly has plenty of exciting potential today. The emergence of silicon architectures for Machine Learning (ML) and Artificial Intelligence (AI) are examples of concepts that have become reality in just a few short years. There are many semiconductor designs currently underway by both new and well-established companies.

Most of the effort has been focused on the data center and the Cloud where the AI functionality currently resides. This will not long remain the norm as the functionality will spread from these two sources and into smaller devices. Given the work that is being done in the AI space, it is likely that processor architectures will continue to be refined, and more processing cycles will be possible with reduced compute resources. We already see discussion and efforts to bring AI functionality to end-point devices. The Amazon Edge, and now Apple’s HomePod and Google Home, are helping to begin the transition into a voice-activated AI interaction. AI functions such as language translation, speech recognition, pattern matching and recognition have been brought out of the Cloud and into millions of homes. This is only the beginning of a growing trend.

In the discussion to move AI to end-point devices, some of the large sticking points are that these devices will not have the power and performance that data centers have. Simple differences in size, access to adequate power resources and different cost models are the main reasons. However, end-point devices by their very nature will not be required to access and process data points on the same scale as data centers. Solutions for IoT or IIoT applications will be required to process far fewer data points to reach meaningful conclusions: essentially the scope of their tasks will be scaled to appropriate levels for the available compute resources. This scaling of the application will enable SoC solutions to contain enough processing power and AI functionality to perform useful and meaningful work for some applications. And a revolution is born.

Taking the concept of ‘relevant scaling’, it is also possible to envision a day when AI applications resident in end-point devices will be created to handle just one task or function. AI for the home could handle all the operations, functions and processes of a residence and be contained in one end-point device. It is still likely to be connected to the Cloud, but most of the processing will be done locally. AI apps could also be similar to the phenomena that captured everyone’s imagination when the iPhone was first introduced. It is not uncommon now to see people with hundreds of apps on their phones. Given the right set of conditions, we could see something similar in specific AI functions in end devices. This would be a new business model for the industry and could create a great deal of additional revenue for systems and software companies.

Semico feels the AI phenomena will prove to be a significant catalyst for growth. If tools are made more powerful, more plentiful, more cost-effective and easier to use, they will be used more often. The emergence of AI will be no different.

So, what does this have to do with the SiP market?

On February 13, 2018, ARM announced the Trillium Project, which offers CPU IP cores created specifically at giving designers the capability of adding ML and AI functions to mobile devices and other end-point devices. MIPS has emerged once again as an independent SIP provider and stated they intend to focus on the AI market as one of their goals. Tensilica and Ceva both have products aimed at ML and AI applications. Other SIP suppliers are following suit with SIP specifically created to facilitate designs for this emerging market.

Added to this is the emergence of eFPGA gates for use in SoCs to create the Convolutional Neural Networks (CNNs) needed to infuse AI functionality into SoCs. In our research, Semico found that eFPGA will be the fastest growing SIP segment with a CAGR of 42.9% through 2022. True, it is starting from a small base, but it won’t remain small for very long, especially as IoT and IIoT solutions grow.

We are already seeing many new non-volatile memory architectures entering the market or in development. Adoption of the right memory IP to enable SoCs for AI applications is one of the critical areas of the SIP market necessary for AI to succeed. Currently Hybrid Memory Cube (HMC) is one of the favored solutions for AI in the data center, but this is a discrete semiconductor solution. End-point devices will need an embeddable alternative that allows the same performance but with a smaller power envelope and footprint. Some of the newer memory architectures like MRAM, STTRAM or ReRAM may provide the right mix of performance and density. More work needs to be done before there is a clear winner.

For the rest of the SIP market-Interface, Interconnect, Security, Logic etc.-the real issue is that most of this SIP is generic, not application-specific. It can be used and applied to almost any end application. Semico believes that some specialization for AI applications will occur in these areas. For example, given the large number of CPU cores per device, does this warrant the creation of an Interconnect technology, Security, Interface, etc. specific to AI? Currently, no such specialization seems necessary from a Cloud-based perspective. Bringing SoC solutions to end-point devices may require another evolution in thinking before these more mobile architectures can be accomplished.

How Big Could the Impact on the SIP Market Be?

An important point is that the IP revenue numbers in the forecast below have an upside if AI is successfully brought to end-point devices. We are already starting to see movement in this direction, but the process will require new types of silicon to provide the right architecture and performance for a successful AI solution. Another reason AI in end-point devices will be a large driver of SIP market revenue is the very large volumes of potential systems that could be created in this market space. While SIP licensing revenues will be strong, SIP royalty revenues will be stronger when they are tied directly to the large number of devices that will be shipping into the market.

Third-Party IP and IP Subsystems Market Revenues Driven by AI Designs

*Forecast
Source: Semico Research Corp., January, 2018

A conservative estimate of the impact of these developments on semiconductor market revenues would be at least a 10% increase in silicon revenues in MPUs, MCUs, Memory, SoCs, FPGAs, Standard Cell, Analog and Sensors. A 10% increase in market revenues, driven by IIoT, SoC, SIP and AI promises to provide a large boost to the overall semiconductor market. As these trends gain momentum, the potential increases will be significant, which is good news for semiconductor and systems manufacturers as well as IP vendors alike.

For more details about Semico’s report, Licensing, Royalty and Service Revenues For 3rd Party SiP, contact Rick Vogelei.