The Automotive Memory Journey: A Busy Road Ahead

Article By : Gary Hilson

Autonomous vehicle rollout is slower than expected as data growth speeds up...

The pandemic has been a speed bump on the road to more advanced transportation as people have been hesitant about ride-sharing and public transportation as well as opting to stay home. Fully autonomous vehicles are moving slower than originally projected, too. However, digitization of the cabin will mean a strong market for memory in the automotive space as cars continue to become servers on wheels.

Kris Baxter

Micron Technology’s Embedded Business Unit general manager Kris Baxter said even though autonomous driving expectations have scaled back, there’s lots of room for growth in the advanced driver-assistance systems (ADAS) segment. Tesla has backed away from autonomous claims and is highlighting ADAS Level 2 and Level 3, he noted.  “We see an increase in that space — adaptive cruise control, lane keeping, automatic braking, and driver monitoring systems. The overall market is going to be very strong in 2021, not for autonomous vehicles, but digitalization of the cabin.”

Longer term, Micron still expects autonomous vehicles to make it into the enterprise-first areas that can afford the costs associated, such as robo-taxis and long-haul trucking applications, but it the meantime, there will still be an explosion of automotive data, said Baxter. “We’ll also see a shift towards centralizing vehicle computing.”

And there’s a lot of computing to be done even beyond onboard infotainment systems as vehicle, autonomous or otherwise, evolve to interact more with the environment and ecosystem of service providers and vendors that provide specific functionalities that contribute that enable autonomous abilities. Aceinna, for example, makes sensing solutions for a number industries, including aerospace and automotive, so it’s not in the computing business per se, but its technology relies on it.

“We’re not doing the computing machines here or anything related to computing power,” said CEO Yang Zhao. “We focus on the sensor.” But there are multiple sensors, including image and vision sensors, that require computing power for processing that data they feed to support navigation and other autonomous vehicle functions. The car of the future, like any robot smart enough to navigate itself, needs to be equipped with a whole host of sensors, algorithms and necessary processing power, he said.

Yang Zhao

A great deal of the computing in vehicles is geared toward working in tandem with external systems such as a GPS satellite for navigation and onboard inertial navigation, said Zhao. “There is a combined algorithm to make the whole system become so fast. You can reach ultimate precision in everything.” However, there’s also the chance the vehicle will get cut off from external guidance systems due to severe weather and other causes. “Your LIDAR can be dead and blinded, but the inertial navigation system inside the car will never get blind. It’s not going to be affected by the outside environment.”

To keep running on their own, these systems require their own computing power and memory, and within the vehicle there also needs to be some segmentation, said Zhao. “The vehicle driving system has to be isolated completely in a closed loop by itself. Nobody else can access that.”

The segmentation of various automotive systems means memory needs will be somewhat varied and draw upon architectures used in other markets, such as mobile and the Internet of Things (IoT), which why all three are part of a single business unit at Western Digital, said Huibert Verhoeven, who is senior VP of the company’s Automotive, Mobile and Emerging Markets Business Unit. “We see quite a bit of overlap historically.” Within the vehicle itself there’s been a transition fully siloed systems with their own memory, such as flash storage for an infotainment system, to a more unified approach. “They’re heading very rapidly towards having clusters of storage for that consumer data.”

Huibert Verhoeven

For infotainment system with maps that are going higher in resolution, he said, there’s an increasing need for an architecture that ties computing and memory together, so that same flash storage expanding in capacity, as is the variety of interfaces, including eMMC and UFS. But if you want a sign that the modern vehicle is becoming a server on wheels, not only are SSDs being incorporated into infotainment systems, Verhoeven said there’s even a demand for NVMe thanks to the “transportation as a services business” where the amount of data that must be logged is tremendous. “They have vehicles driving around that truly are data centers on wheels.”

While some functions of autonomous vehicles involve real-time with communications systems, Verhoeven said there also scenarios where they are accumulating terabytes of data over the course of several days for the purposes of fleet management, for example. “Many of our customers actually choose to at some point to pull those drives out and to load them in racks and to offload the data that way.” A removable form factor and serviceability becomes important, he said.  “They like the fact that they can easily operate to high capacities, and if something were to go wrong with a drive, they can easily replace those products.”

Many applications are modern vehicles are leveraging core technology for other use cases in industrial, mobile and server scenarios, said Verhoeven. NVMe is appealing in some cases because of the potential data volume and latency requirements that are not unlike what you’d find in a data center. But while there is some logic in reducing the number of separate data siloes with the vehicle, there still needs to some segmentation—there would be risks in putting mission critical ADAS features on the same data storage media as onboard entertainment for the kids, he said, with certain subsystems accessing data from a common pool in that cluster as it balances cost and efficiency with safety and reliability.

The level of autonomy is also informing how much memory and storage ends up in a vehicle, said Verhoeven, with Level 4 and 5 expected to require exabytes by 2027. “But the vast majority actually is driven by things like infotainment and ADAS that are very much a Level 2 and 3 discussion.” Western Digital is seeing its business grow as OEMs embark on the more advanced Level 4 and 5 implementations, however, which is guiding how the partition of different data systems evolving.

Even a vehicle that’s not fully autonomous requires a great deal of onboard data storage as the cabin further digitized and designers add multiple, high resolutions screens infotainment consumption (Courtesy Western Digital)

But even just at Level 2 and 3 there’s are so many subsystems that produce information has to be displayed as part of larger sub clusters, said Verhoeven, even in lower end vehicles with some ADAS features. Even without self-driving capabilities, there could potentially be terabytes of data distributed throughout the car. “What we are seeing now is that all the mainstream OEMs are putting the larger displays into a car, but also multiple displays into the vehicle,” he said. And as advanced features trickle down to become standard in lower end vehicles, memory and storage requirements will rise, which means it won’t take long to need at least half a terabyte.

Increased autonomy and cockpit digitization is driving a transition from a fragmented architecture with individual (ECUs) around the car to having a single domain controller, said Micron’s Baxter. Automotive engineers will increasingly turn to multi-chip packages, which combine memory and storage into one streamlined package, and GDDR6 will replace low-power DRAM for some applications.

There has been a profound shift for the memory industry as the automobile has become a major driver of memory technologies, he said, and it’s required changes to how memory is defined and designed to get aligned with the market timing and requirements of the auto industry. At the same time, established technologies such as NOR flash still have a role in in supporting fast boot capabilities to meet driver expectations for instant-on when they turn the key in the ignition.

Power is a big challenge because delivering data center on wheels requires a great dealing of computing well above 600 tera operations per second (TOP), which can demand as much as 1,000 watts of power. It’s not just limited to the processors, but also the memory. Power consumption becomes even more critical for fully electric vehicles and creates a thermal management challenge for automotive engineers.

Micron’s approach is to look at where power/performance tradeoffs are possible by employing low-power, energy-efficient memory such as LPDDR4X. More advanced processed modes, meanwhile, come with commensurate improvements in power efficiency, which are incredibly crucial for autonomous cars that run power-hungry high-performance, AI-enabled applications.

Complexity is also a challenge, and underestimating it is why Level 4 and 5 autonomous vehicles have been slower out of the gate, said Robert Bielby, Micron’s senior director for automotive system architecture. “There’s still quite a bit of a disparity as far as what is really the amount of compute performance that’s required to realize autonomous driving.”

automotive memory demand
Based on Yole Development’s Rapid Adoption scenario, demand for automotive flash storage is growing, with NAND auto bid demand expected surge for autonomous Level 3 to 5 vehicles beyond 2022 (Courtesy Western Digital) Click on the image for a larger view.

Like Western Digital, Micron sees a move toward clustering that would see mission critical data segmented from onboard entertainment content, while also leveraging common pools of data for efficiency, such as a central storage point where all maps are in a single location but used for different applications and this further emulates how a server is architected, said Bielby. “This is where you’re going to start to see the concept of virtualization, so I can access the common storage pool from two different applications.”

There’s increasingly more attention being paid to security and safety, he added. “What’s important to the industry is that any and all memory that’s being used in the automobile is auto qualified.” And that’s a wide spectrum of memory products, whether it’s ADAS that’s driving DRAM uptake or GDDR6 to support multiple screens in vehicles, said Bielby. “There’s been a move towards NVMe and SSDs as a whole.”

A lot of designers are taking a step back from the previous approach of adding more functionality and more capability in a piecemeal manner. There’s a shift to taking a more holistic view to toward the architecture. “Rather than have multiple discreet eMMCs or UFS devices onto the system, there’s a trend towards moving towards more of a consolidated, centralized storage approach.” Anything that goes into a vehicle must automotive grade, he said, so while emerging memories may have some use cases, “there’s the grease that’s required to get there. My expectation is that the technologies will mature and become available for these applications.”

Tom Coughlin

Tom Coughlin, president of Coughlin Associates, said automakers are just starting to move to SSDs for infotainment data storage as prices are coming down to a point where it makes economic sense. And because there’s so much memory needed to build a server on wheels, it’s inevitable that emerging memories will be considered, but they must be able to handle the extreme environments associated with automotive use cases. MRAM, for example, has been shown to handle extreme temperatures and could potentially replace NOR flash in some applications, he said.

In general, there’s a lot of tradeoffs to be considered when selecting memory devices for automotive environments, said Coughlin, which means finding a balance among capacity, endurance and retention. But memory also has to work with a lot of other evolving technologies going into cars, including communications and interface protocols such as Ethernet and TCP that comprise the networking aspect of a server on wheels, all of which support external data traffic to edge environment where additional processing and AI takes place.

The memory and compute in vehicles and the surrounding infrastructure will need to keep up with the 5G speeds that are believed necessary to communicate with cars in real time, as well as all the training that needs to happen, said Bielby. “There are millions, if not billions of miles of data that you need to collect from road data that’s going to help you build or train a model that’s going to be capable of driving in complex automotive environments.” The training side of the equation relies on high performance data centers, already hungry for memory, and connectivity in general is going to be a key component of autonomous driving, he said, to deliver real-time updates such as traffic conditions, as well as opportunities for algorithm improvement.

Over the longer term, 5G will play a role in realizing cellular V2X, which means not only are cars talking to the surrounding infrastructure, but also to each other, which essentially makes vehicles on the road become a cluster of servers computing together without necessarily requiring a central cloud. “That’s going to allow the ability to predict what’s happening in the environment and understand what the behavior of a vehicle is five or 10 cars in front of me.”

Subscribe to Newsletter

Leave a comment