With industry titans like Nvidia, Intel and ARM jumping on the bandwagon of custom silicon, it can be safely concluded that the age of custom silicon is well upon us. This has interesting repercussions for the semicon industry and changes how chips are designed and sold. The clear winners are the foundries. Other players like IP companies, fabless players and design services entities can all thrive, if they find the right Product Market Fit and execute well. The market is huge and growing, so there is no shortage of opportunities for products or business models.
The Need for Custom Silicon
A variety of factors are driving the market towards custom silicon. The initial impetus came from the demise of Moore’s law and the end of free lunch. You can no longer get increased performance or better area just by hopping a node or two. Power does still seem to improve with node jumps. But that is not just due to node shrink. It is also due to new transistor types, better materials and techniques like backside power delivery. The consequence is that custom silicon is mandatoryif you want an optimal design. This is very feasible since most fabless parts target the majority requirements of the market and cannot afford to have variants targeting niches. So, any large volume customer who wants maximum optimization has no option but to go the custom way. Hyperscalers and the cloud giants started the trend, followed by auto companies for the ADAS segment. Power-sensitive edge applications from sensor nodes to edge ML devices would be next. But other applications like storage controllers, interconnect parts, chiplet components, and mixed-signal parts are also not far behind.
The second reason for the trend is the arrival of Chiplets. Chiplets inherently lend themselves to customization since only a max of 1-2 chiplets need to be optimized in a chiplet based design with 6-8 chiplets. The customization effort is significantly reduced and the design effort democratized in the process. Companies that never would have considered a custom part are now considering them as a viable option. This is by no means restricted to high end silicon. Automotive and industrial parts are also amenable to chiplet-based optimization, especially when the customization is needed for mainly IO, RF or analog components. It also reduces lifecycle costs for sectors like high end transportation, medical and aerospace markets, since mid-life upgrade via chiplet replacement can result in reduced compliance recertification efforts.
The third reason is that the advent of high-speed memory like HBM and high-speed interconnects require custom designs to take advantage of the latest standards. It is simply not viable for a fabless vendor to cater to every interconnect or memory standard with all their variants. CXL support in x86 server parts is a case in point. This requirement is partly met by chiplets, but the IO die still has to be customized. IBM’s open memory interconnect is an interesting approach to solving the rapidly evolving memory standard problem. It had the added bonus of a free L4 cache. High-end server parts have a much longer life cycle than memory standards and starting from the midpoint of their lifecycle, tend to have slower memory interfaces compared to the norm in the industry.
RF devices also end up being custom parts when max efficiency is required, necessitating custom RF circuits
Industry Strategies to address this trend
The strategies can be broadly classified based on whether the strategy is IP centric or service centric . IP led services or Services bolstered by IP offerings. Both meet somewhere in the middle. But company culture, R&D investments, ownership of IP, ODM offerings all combine to determine gross margins and place in the pecking order. It is intuitive to think companies that offer the most comprehensive set of services and IP will dominate. While that is true to an extent, excessive integration may also lead to bloat and inefficiencies. There is a goldilocks zone in this business waiting to be discovered! The right combination of internal capabilities and partner capabilities matter. Chiplets naturally drive the trend towards an eco-system approach since they exemplify the best of breed approach, but larger companies or IP-centric companies will I think end up dominating the market for standard chiplets. This in turn may end up commoditizing standard chiplets!
Fabless companies and processor IP companies will lead the trend since they are used to driving SoC architecture. While they will suffer a reduction in gross margins, it is not really as bad as some fear since custom design offerings can lead to steadier margins and can significantly reduce market risk. This was cited by Intel recently in the context of custom Xeon offerings. Think in terms of an investment portfolio. While a 100% small cap portfolio can deliver excellent returns in certain phases of the market, the ride is not smooth by any definition! A diversified portfolio works better, both in investments and in silicon offerings. Everyone looks at the poster children of fabless and IP business while ignoring other players who do not have stellar returns. Custom silicon possibly (just possibly!) may even the playing field for smaller fabless companies.
Design services companies immediately benefit since their gross margins will go up. But there is a big IF! Only those which have comprehensive chip architecture capabilities, end-to-end service offerings and standard platforms can cash in on this trend. The others will definitely benefit from this secular trend since chip design starts are increasing but only a full custom offering will lead to better margins.
Packaging tech or access to it will also determine winners in the higher end of the segment. This was typically thought of as the domain of the OSATs or fabs. But with 3D interposers becoming a core part of the chip architecture, players in the highest end of the market need to have packaging chops, in-house custom SEDRES and other critical IP.
Niche segment players in the security domain and functional safety also need to leverage the move towards custom silicon. I am sure other niches will open, creating space for a lot of custom chip startups.
To conclude, the chip industry seems to be heading towards a broad vision of custom silicon. But the paths are myriad (and so is the pain!) and so forecasting beyond a point is futile. Increased chip starts are a given. Those who execute well will reap the rewards.
