Recently there was a flurry of news around Microsoft starting to build its own custom AI chip in-house. Code named Athena, these ARM based chips are being planned to reduce the tech giant’s dependency on NVIDIA for its highly sought after GPUs. This follows a long trend of tech giants investing in designing custom chips for specific use-cases, thereby reducing their dependence on the likes of Intel and AMD. These recent shifts made me consider spending some time on how the automotive industry has fared in the custom SoC drift.
What do we mean when we say ‘Custom Chips’
Let’s get some alignment on the terminology out of our way, first.
An SoC (System on Chip) is a silicon chip that contains one or more processing cores, along with on-chip memory, peripherals, and various controllers. Because an SoC contains a microprocessor, as well as the other necessary components, a custom SoC greatly reduces the component count, so that the resulting circuit board is much smaller.
SoCs/ASICs (Application Specific Integrated Circuits) are terms used rather interchangeably. Very simply, these are silicon chips where the designer implements components for compute, memory/storage, power management and other critical functions based on specific needs for the designer. The latter point sets customer SoCs apart from general purpose chips manufactured by traditional companies like Intel, AMD, Qualcomm and so on.
What are the designs being optimized for?
Custom SoCs serve a variety of needs within a vehicle, including:
Supporting control units - ECUs etc.
Navigation, telematics, V2X purposes
Sensor control & Sensor fusion - Camera (Vision), Radar, Lidar
Media / Infotainment
Auxiliary component control including HVAC
Safety critical applications
Energy / battery management
You can clearly see that there are a vast number of use-cases within a vehicle, where software plays a critical role and the microcontroller (and needless to say the embedded software that runs on it) becomes central to this whole ecosystem. But peel the layers a little, and there are mitigating factors that arise, when designing SoCs for all of these use-cases. When designing custom SoCs, these are some key factors and trade-offs that designers need to grapple with:
Throughput vs Latency is a key question that has implications on memory architecture and how these components are interconnected. Each use-case has different quality of service standards, and these trade-offs need to be balanced.
When the number of subsystems to be supported by a single chipset increases, the scale of complexity increases further.
It is possible to replace discrete components of the chip with software functions (for example, share a RAM buffer with multiple processing nodes, by controlling access through software), but this results in increase in software complexity
Software complexity can increase the probability of faults & errors, and also increases the vulnerable surface for security threats
Any custom SoC design must also comply with ISO-26262 road vehicle functional safety standards
Along with all of the above considerations, any automotive subsystem has additional physical constraints including dimensions, vehicle fitment and operating conditions (including vibration, radiation, temperature etc.)
So why are custom SoCs still a big deal?
When there are these many trade-offs, why do automotive OEM majors and the big-league tier-1s spend so much on custom SoCs? This neat explainer by Nick Fiorenzi sums it up really well:
When dealing with serious volumes, it makes sense for users who have specific applications and in any case throw in serious investments in vehicle systems development to take control of SoC design. OEMs follow a combined approach of specialist in-house teams (which usually own the spec) and leverage external consultants or specialist organisations for downstream activities like prototyping, quality assurance, functional safety testing, certifications etc. This gives the OEMs and Tier-1s the leverage to work directly with semiconductor fab companies to fulfil their needs. Along with gains like increased autonomy, and better performance for specific needs, designing custom SoCs in-house also gives OEMs a valuable degree of supply chain de-risking.
Where are these trends headed next?
Since the time Volkswagen introduced transistors to control fuel injection in 1968, the usage of semiconductors in vehicles has seen several structural shifts. The gradual progress towards L3/L4 vehicle autonomy, improvements in connectivity and the rising share of electric powertrains are shifting existing constraints on SoC design, as well as bringing in additional use cases. EVs lend themselves to simplification of the overall vehicle electronics architecture; improved 5G connectivity implies more computing & processing can be shifted off-board to the edge or cloud based systems. Headwinds like 5G rollouts and uptick in value created from vehicle connectivity means that, according to McKinsey, the automotive computing ecosystem will possibly coalesce across three dominant archetypes.
The automotive SoC market, which accounts for 20% of the global SoC industry, is set to grow at about 7.8% CAGR to $26.8B by 2028. Apart from vehicle autonomy, improvements in cockpit & infotainment functionality and these functional designs being brought increasingly in-house are the key drivers for this growth. Progress on 5nm process nodes adoption for automotive SoCs, improvements in in-vehicle networks and miniaturization of ADAS components will be the key trends shaping custom SoC trends for the automotive industry.