Li Weili from Alibaba DAMO Academy: R&D for Automotive RISC-V Functional Safety Requires 3 to 5 Years

Gasgoo Munich – Automotive Semiconductor Ecosystem Conference 2026

On September 16, Li Weili, a processor architect at Alibaba DAMO Academy, addressed the advancements and challenges surrounding RISC-V in automotive electronics. He highlighted that RISC-V applications have matured in both China and internationally across four dimensions—IP, IC, complete systems, and software ecosystems. However, he emphasized that the development of functional safety remains challenging, particularly due to lengthy processes and difficulties in achieving diagnostic coverage targets for ASIL-B software.

Li pointed out that developing functional safety products necessitates an iterative process involving both product and process certification, which typically spans three to five years. For ASIL-D systems, a Dual-Core Lockstep (DCLS) approach is often employed, which, while straightforward in logic, demands twice the area and power consumption. In contrast, ASIL-B systems use Software Test Library (STL) diagnostics to maximize power, performance, and area (PPA), requiring 90% diagnostic coverage with only a 10% to 20% increase in area—a notably significant challenge.

He outlined three primary constraints in STL development:

1. Runtime: STL execution must not disrupt current functions, leading to code segments being restricted to a few thousand cycles to ensure interrupt response latency remains unaffected.

2. Storage: Due to strict memory limits in automotive applications, STL must be compact, which restricts the variety of instructions and test cases.

3. Access Safety: Interacting with Device-type I/O hardware can lead to irreversible states, further constraining software diagnostics.

Li also highlighted a critical gap in diagnostic coverage between Register-Transfer Level (RTL) simulation and netlist simulation. The discrepancies can be over 5%, potentially reaching 10%, suggesting that achieving 90% coverage at the RTL level may not suffice for netlist validation, compelling teams to aim for 95% or higher during RTL design. He noted difficulties in covering internal MCU modules and challenges posed by bus fault injections that compromise testing efficacy.

Looking to the future, Li urged more vendors to participate in the RISC-V automotive ecosystem, aiming to enhance MCU-level IP for ASIL-B and ASIL-D standards and to create high-performance computing cores for autonomous driving. The aspiration is to foster a richer, more resilient ecosystem in the automotive sector.