ARM is extending its work on ISO 26262 safety packages for automotive systems beyond the Cortex-R devices supported in a documentation release earlier this year.
Recognizing that advanced driver automation system (ADAS) designs that become increasingly autonomous will need to use a range of high-speed processors, ARM plans to provide safety documentation for its Cortex-A processors in the short term and extend that out to the Mali GPUs, which are likely to be used in applications such as 360° view generation.
Richard York, vice president of embedded marketing at ARM, said: “This is the next stage of what we are doing for safe and reliable systems. We found we could apply our expertise gained so far to the A-series processors, including the new A72. This is the first time that anyone has done this for high-end CPUs.”
ARM is working on the basis that ADAS requirements will lead to a massive increase in compute horsepower. “All these applications have a functional-safety component to them,” York said. “We are expecting a hundred-times increase in the amount of compute in a vehicle over the next decade. A hundred times is probably quite a conservative number. Once you start putting compute engines for vision, it spirals up quite rapidly.
“However, automotive systems end up having some quite challenging thermal constraints. That will keep the amount of compute limited. There is a really delicate balance and tradeoff over what is practical and realistic.
The different architectures of the devices will lead to a stratification of safety certification within the systems, York said, such that the advanced processors will have fewer specific safety features than cores such as the Cortex-R5. “When you get this amount of compute you just can’t afford to do lock-step [dual-redundant] execution. You can’t afford the power or the silicon area. There needs to be a safety hierarchy.
York added: “Things that interpret the world will have a slightly lower level of safety. Those that make control decisions will have a higher level. The closer you get to the final decision, the more safety-critical it is. But it’s not a straightforward problem. We want to do some of this work to make the systems design more straightforward for customers.
Deep learning trend
“The challenge we are trying to tackle is what is an appropriate level of safety for these big applications processors. GPUs will be a part of this story in the future. A lot of these systems are about processing vision and some of these GPUs are very good at that.”
The development of more sophisticated systems based on deep-learning techniques may feed into architectural modifications and new processors. “Deep learning is an area of real interest for many companies. Our R&D team is looking at it like everyone else. What exact form of compute will apply is still an open question,” York said. “There may be new forms of compute engine. However it evolves, the basics of how we create safety packages and failure-management reporting features we can still apply to to that. This safety-oriented work has become a basic part of how we design our products.”