Data Storage Issues Grow For Cars

Autonomous vehicles will generate huge quantities of data. Where will it be kept?

popularity

Adding safety features into cars and making them increasingly autonomous are rapidly creating a big data problem. More sensors produce more data, which has to be processed, moved, and ultimately stored somewhere in those vehicles.

self driveing electronic computer car on road, 3d illustration

Exactly how that will be achieved isn’t quite clear yet. However, there is plenty of discussion on that topic—and for good reason. A new 2017 car will generate about 20GB of data every day, assuming the car has 2 cameras, 16 sensors and is driven 1 to 2 hours a day, said Claudio M. Camacho, head of marketing for Tuxera.

Wally Rhines, chairman and CEO of Mentor Graphics, believes that autonomous vehicles ultimately will require a hybrid approach of distributed and centralized memory and storage.

“In the central ECU, you will need a really big memory,” Rhines said. “You will have to do fusion, which will require main memory. With visualization, there will be multiple cameras and you have to synchronize everything. If you have several different cameras looking at the same image, you don’t want to corrupt the data so at least some of that will be raw video. That will require huge bandwidth. You can’t get enough of it, because if you’re driving down the Autobahn at 150 miles per hour, how do you recognize the image of a dog in time to avoid it? You can’t solve that if the data doesn’t get to a central CPU, and you will need a lot of memory to do that. Some of it will have to be centralized. Other pieces that are less critical can be distributed.”

Next-generation cars will include 20+ computers, at least 10 of which have storage, Camacho noted. “The current trend is to keep all storage separated, so each box (i.e., each Linux computer) has its own storage. Storage sizes vary from 8GB to 256GB. Some of the most important boxes are black-boxes (for event recording of self-driving cars), V2X (communications from Vehicle to Anything), ADAS, cluster, IVI (in-vehicle infotainment) and EDR (event data recorder). The telematics box also has storage to buffer the over-the-air updates coming through the 5G network. That one uses around 32GB of storage.”

Storing data in cars isn’t an entirely new concept. “Storage is a fairly wide term for automotive, and almost every MCU — there can be 100 different MCUs in the car — typically has some sort of NOR flash for program code,” said Ron DiGiuseppe, senior strategic marketing manager for Synopsys’ Solutions. “That’s not changing. MCUs have boot ROM and program flash controllers, and they are always going to be there.”

For infotainment systems, meanwhile, various types of DDR storage have been implemented because of the graphics processors. “There have been hard drives that have been storing song libraries, and have been in various entertainment systems,” DiGiuseppe said. “Hard drives now tend to migrate toward SSDs due to the high vibration environment. Spinning disks may not always meet the reliability that SSDs do for infotainment storage.”

But the move toward Advanced Driver Assistance Systems (ADAS) is something of a game changer for data storage. For one thing, it will depend on much higher-performance processing, and that in turn requires a whole different set of memory architectures.

There are number of factors impacted by ADAS. First, a lot more functionality is being added to those SoCs, which are now going into much more high-end performance requirements. Until now, processors in cars might have included a 300MHz embedded microcontroller. ADAS processors use multiple 64-bit multi-core processors from ARM and others, some developed using the latest finFET processes. These devices have the same kinds of storage needs you would expect to find in other high-speed computers, including LP-DDR4 running at speeds of up to 3,200 Mbps, which is considered the sweet spot for a lot of newer ADAS processors.

“It’s a rather new phenomenon for automotive applications to have such advanced processors,” DiGiuseppe said. “64-bit processors in a car is new, and that started with a lot of the ADAS vision processors that need 64-bit operation. These also need a vision co-processor, like an embedded vision co-processor sitting next to the host processor, and this is also running at pretty fast data rates with very advanced vision algorithms. When you throw in the requirement of vision co-processing with 64-bit host processing, the external memory requirements are growing in terms of density and performance.”

And this is just the beginning. “We’re starting to see deep learning in automotive,” said Jeff Bier, founder of the Embedded Vision Alliance. “There is an enormous investment there now.”

All of this is enabled by economies of scale in sensors, processors, software development tools and techniques, algorithms and a commensurate improvement in engineering skills, Bier said. “The rate of improvement in processing and performance per watt is increasing. People are seeing the market for this and they’re putting money into it. If you look at algorithms before deep learning, they were improving rapidly. Now, with deep learning, the growth curves are increasing rapidly.”

All of that means more data to process and store. In addition to the DRAM increases that will be required, because of new ADAS applications along with traditional infotainment, there is further use of NAND flash for storage. This would be in support of mobile storage standards like embedded multimedia card (e.MMC), and maybe moving to universal flash storage (UFS) or solid state drives to address the requirement for storage to continuously update real time maps when all the automakers start rolling out their autonomous driving offerings.

“It’s well acknowledged that mechanisms for autonomous driving include real-time vision that’s occurring as you go,” said DiGiuseppe. “That’s in addition to existing maps and real-time map updates, where the car can compare for trajectory planning such as what they’re going to do next, and even for anticipating awareness of surroundings like, ‘Will this pedestrian step off into the street?’ In trajectory prediction and some of the application awareness, real-time maps are required, and that’s going to take more flash storage. So there will be an increase in the amount of NAND flash on ADAS processors to accommodate the real time map updates.”

Architecting memory, storage
Hubert Bailey, director of storage solutions applications engineering at Marvell, pointed to other areas where storage has been used within cars, such as GPS systems, entertainment systems, web access, and logging. GPS systems use storage to store the map set for the car, and then rely on the GPS antenna to determine the location of the car on the maps, while entertainment systems have used storage for storing movies and audio to be played in the car. In some cases, the movies use a DVD player connected to the entertainment system, and audio systems have been savings CDs or downloaded MP3 files to internal storage for playback. Further, as web access increases in the car, the need for caching goes up. This is another place where storage can have some impact.

But when it comes to architecting the system to place the memory and storage in the optimum location, it will depend on the system where the storage is located. Most entertainment and GPS systems currently do not have a built-in storage port, and Marvell devices have been used to augment the system to add SATA storage ports.

As far as whether the memory and storage will be shared, Bailey explained that different systems typically have operated independently. But as the level of integration goes up, the storage may become shared. Another option would be distributed memory and storage, which Bailey sees as an area in question, and Tiered suppliers are awaiting direction for the automotive OEMs.

Specifically, for ADAS ECUs, these generally have one application such as automatic emergency braking, each of which typically has its own ECU and its own ADAS application processor associated with it, DiGiuseppe observed. “For that application, you’re going to need between 4 GB and 8 GB of memory in LP-DDR4. There are multiple ADAS ECUs, each with their own host processor, with their own separate memory.”

In keeping with the integration trend in the general semiconductor space, the automotive industry is following suit and starting to integrate some of those separate ADAS ECUs. “This is because the SoCs are able to integrate those applications into centralized ADAS processors as opposed to one ADAS ECU with its own ADAS processor per application, there’s no reason why you can’t architect so there are multiple applications in one ADAS processor,” DiGiuseppe added. This will also increase the memory requirements depending on the applications that are combined into one processor.

Conclusion
No one is quite sure how much data ultimately will need to be stored in a car, for how long, or in what form. Initial concepts for autonomous vehicles involved much more processing in the cloud, and target dates for introducing autonomous vehicles into the market as of just several years ago were in the 2030 timeframe. Now Ford has announced it will have autonomous vehicles on the road within the next four years, and Tesla already has a self-driving mode working inside its cars.

That has stepped up requirements for storage and advanced processing architectures, and many of those architectures are works in progress. But there is no question that they all rely on moving, processing and storing much more data than in the past, probably including at least some RAW image and video data. Some will need to be recalled quickly from memory, some will need to be stored for different periods of time using some type of storage. And nearly all of it will have to be architected into the system and managed in a way that can be utilized by a fast-moving safety-critical system.

These are difficult parameters for designers, made even tougher by extreme conditions that can affect reliability and unknowns that seem to change dramatically by the car model year.

Related Stories
Prioritizing Vehicle Data Traffic
Challenges grow for classifying and tagging huge volumes of data from connected cars.
Way Too Much Data
Each new node and architectural change results in an explosion of data, making optimization of designs significantly harder.
What Can Go Wrong In Automotive (Part 3)
Why power has become so important in car electronics; the challenges in making autonomous vehicles reliable enough; adding margin for safe modes of operation.