Marvell recently introduced custom silicon solutions targeting the growing market for data processing and AI accelerated solutions for the data center and automotive markets.

AI processing (or inferencing) operations are outstripping conventional processor compute requirements. Marvell notes that the computational requirements for inferencing has grown by 50x every 18 months. In comparison, it has taken Moore’s Law scaling of transistors/die 120 months to scale at the same rate. Marvell sees the computational capacity for AI silicon doubling every 3.5 months, compared with logic density doubling every 18 months and doubling of SRAM density taking more than 48 months.

Marvell believes this gap between traditional general purpose processors and the requirements for AI accelerator solutions is driving a growing requirement for custom solutions that can provide an optimized blend of energy consumption and performance. Marvell is leveraging over 25 years of custom silicon design experience to meet this need. The company’s ASIC business traces its roots to GlobalFoundries’ 2015 acquisition of IBM Microelectronics. The ASIC business was then spun out as a separate standalone business, Avera Semiconductor, in 2018 before being acquired by Marvell in 2019 to become Marvell’s ASIC Business Unit. The BU boasts over 2000 ASIC designs, 14 process nodes, advanced packaging technologies, related intellectual property and a range of related design services.

The company’s latest best of breed offerings are based on 5nm process nodes that can offer a reduction in both chip die area and power of over 30%, compared to Marvell’s 7nm ASIC technology. Common custom solutions include 64-bit Arm® multi-core processors combined with high speed interfaces, security processing, machine learning and data processing accelerators. Memory requirements are underpinned by high density SRAM and Marvell has also collaborated with Micron on the first 3D stacked DRAM over a logic die Hybrid Memory Cube (HMC). Advanced packaging technologies include 2.5D and 3D packaging as well as MCM (multichip module) concepts using chiplets. The company uses TSMC’s bulk FinFET process for the most advanced 7nm and smaller ASICs as well as using GlobalFoundries for SOI and older process technology requirements.

The company combines this capability with a range of business models designed to offer customers a number of flexible ways to work with Marvell’s offerings. This ranges from working with COPD (customer owned physical design) with Marvell providing the layout to full turnkey solutions that encompasses everything from the design architecture to tapeout, providing support that includes software and firmware.

Marvell is engaged with a number of data center and automotive manufacturers to develop custom AI silicon enabling high performance, low power solutions for specific workloads.  This includes advanced machine learning solution providers like Groq, one of the companies using Marvell’s latest custom ASIC platform to underpin what will be reportedly the industry’s first peta OP/s AI accelerator-on-a-chip.

For the automotive sector, Marvell is already a leading player in networking and storage capabilities, and the ability to bring together secured interconnect, storage and compute is designed to position Marvell as a one-stop shop for the automotive industry. The company’s automotive offering looks to incorporate Marvell’s standard product and turnkey capabilities based around the company’s full suite of standard IP building blocks, combined with custom silicon to underpin the moves towards centralized architectures that will underpin future electronic platforms and enable the move towards greater levels of autonomy.                      

Marvell’s confidence in its ability to support automotive reliability requirements is underpinned by AEC-Q100 qualified parts and a holistic approach that encompasses an understanding of intrinsic failure mechanisms, fundamental IP that underpins a robust chip design methodology and procedures in place to incorporate packaging and test. This holistic approach is further enhanced for aerospace and automotive industries to realise less than 1 DPPM (defect parts per million) for the company’s offerings.

Marvell is seeing automotive industry demand for its custom silicon on two fronts:

• Robotaxi and self-driving solutions are leading the charge on autonomy and provide the primary opportunity for custom ASICs with a need to support AI operations requiring 100s of TOPS of performance.
• Traditional OEMs, including both legacy and emerging players, are looking to incorporate domain- and zonal-based centralized architectures, with compute capabilities focused on autonomy.

Marvell’s offerings are designed to offer OEMs a means to differentiate the compute capabilities of their platforms while offering smaller dies and a smaller cost per piece, compared to standard SoC/GPU offerings. Furthermore, the ability to incorporate Marvell’s ASIC technology using advanced MCM packaging that can incorporate multiple chiplets allows OEMs to create a scalable autonomy roadmap ranging from Level 2 to Level 5. Marvell is presently engaged with 24 OEMs, including seven out of the top ten, to serve their automotive in-vehicle-network requirements for current and future car platforms.

Now, check out the report “Gateway Module Market Outlook 2018 - 2027: Evolving Vehicle Architectures to Dictate Performance Focused Requirements” (subscription required) for some related analysis and thanks for reading!

Feel free to contact me to discuss this post and the underlying questions raised. For more information on Strategy Analytics’ extensive coverage of the automotive industry, take a look at the PBCS (Powertrain, Body, Chassis & Safety), AVS (Autonomous Vehicles Service), AIT (Automotive Infotainment and Telematics) and ACM (Automotive Connected Mobility) services.