Inside Mobileye's EyeQ Ultra Chip for Future Self-Driving Cars
Check out our CES 2022 coverage.
At the Consumer Electronics Show, Mobileye revealed the next generation of its flagship system-on-a-chip (SoC), called the EyeQ Ultra, that is purpose-built for cars that can almost completely drive themselves.
Unveiled at CES 2022, the EyeQ Ultra aims to provide enough processing power for what experts call "Level 4" autonomous driving, which means that cars can operate in self-driving mode in most, but not all, situations.
Intel’s autonomous driving unit has become an early leader in locking down the brains of self-driving cars. For years, it has offered camera-based systems with a package of its hardware and software to auto makers trying to add safety features such as emergency brake assist and lane-departure and blind-spot warnings—also called advanced driver assistance (ADAS) systems—to cars. There are millions of those systems on the road today.
The company has previously said that it has shipped more than 100 million EyeQ processors to date.
Mobileye has promoted its ADAS systems as a stepping stone for self-driving cars. Last year, the company announced that it started testing prototypes of self-driving cars in several major cities, including New York. The autonomous driving giant is also developing in-house lidar sensors to complement the radar systems it is also building in-house. Paired with its cameras, the sensors will help vehicles “see” more of their surroundings.
With EyeQ Ultra, it is upping the ante in its battle with Nvidia and Qualcomm to power autonomous cars in the coming decade. Mobileye already sells a package of hardware and software that can add a Level 4 self-driving mode to cars, but it uses several chips that are too power-hungry for cars suited for consumers. Integrating such a system in a single SoC will improve its cost and energy efficiency, which also allows it to run cooler.
Under the Hood
Tapping 5-nm process technology takes the performance of the EyeQ Ultra chip to another level.
The EyeQ Ultra can supply up to 176 trillion operations per second, or TOPS, to carry out artificial intelligence workloads in cars and up to 4.2 TFLOPS to handle more general-purpose processing jobs, the company said.
Mobileye said the EyeQ Ultra pumps out performance that equals 10 of its latest EyeQ 5 chips combined in a very power efficient envelope of less than 100 W. The EyeQ 5 is set to be used in production vehicles in 2022.
While not on the same level as Nvidia’s Orin and Qualcomm’s Snapdragon Ride chips in terms of performance, Mobileye CEO Amnon Shashua said the Ultra chip has more than enough muscle to control a consumer-grade autonomous vehicle. It also wins on power efficiency, he claimed. Reducing power is growing in importance for electric vehicles as self-driving systems will have to run off the same battery as the powertrain.
"It is not all about TOPS," Shashua said during a press conference at CES this month. "The efficiency is born out of understanding the very, very tight interplay between hardware and software.”
Mobileye is also keeping costs in check by cramming more performance in a single monolithic chip. The Ultra will be priced at less than $1,000 when it is widely available to auto makers, Mobileye said.
Mobileye said prototypes of the Ultra will be ready in 2023, with full automotive-grade production set in 2025.
Hitting the Accelerators
The EyeQ Ultra contains 12 dual-threaded CPU cores based on the open-source RISC-V architecture, moving away from the MIPS CPU cores in its predecessors. Also under the hood are a graphics processing unit (GPU) from Arm that can run up to 256 billion FLOPS and a vision processing unit (VPU) to help cars "see" the world around them. The chip integrates an image signal processor (ISP) and a set of H.264/5 video encoding cores.
The chip is supplemented by four classes of proprietary accelerators, each tailored for a specific set of tasks. The chip includes 64 such cores, including 16 convolutional neural network (CNN) accelerators.
There are also 48 general-purpose accelerator cores, including eight programmable cores based on its coarse-grained reconfigurable architecture (CGRA); 16 cores based on its hybrid very-long instruction word (VLIW) and single-instruction, multiple-data (SIMD) architecture; as well as 24 "barrel-threaded" CPU cores.
The chip adds internal redundancies for safety as well as a security subsystem. The EyeQ Ultra can be paired with a separate microcontroller (MCU) for ASIL-D systems, Mobileye said. It also supports LPDDR5X memory.
Mobileye said Ultra can process data from two sensor subsystems–one camera-only system, and the other combining radar and lidar–and the vehicle's central computing system, high-definition map, and driving policy software. All the cameras, radar, and lidar sensors give the vehicle 360-degree coverage of the world around it.
The EyeQ Ultra is the "crowning jewel" of its product line, Shashua said, likening it to an "autonomous-vehicle-on-a-chip."
While it previously paired its chips with its homegrown software, sensors, and apps in a "black box", Mobileye said that it uncoupled the software and hardware in the EyeQ Ultra chip to give its customers more flexibility. Mobileye said Ultra is not required to run its in-house software, which allows auto giants to buy the hardware or software à la carte and run their in-house apps on top of them—or buy the full set from Mobileye.
Eyes on the Road
At CES, the automotive chip giant also announced a pair of 7-nm processors for ADAS: the EyeQ 6L and EyeQ 6H.
The EyeQ 6L is ideal for the front-facing camera systems in cars. The chip is faster and less power-hungry than its predecessor, the EyeQ 4, meaning that it can be mounted behind a car's windshield to watch the road for dangers. The chip, which fits inside a 50% smaller package than its predecessor, brings more performance to run advanced camera-based safety features such as automatic emergency braking (AEB) and lane control.
Mobileye started sampling the EyeQ 6L last year, and it is set to enter production by the middle of 2023.
The EyeQ 6H is targeted at more advanced safety features in cars that rely on several surround cameras, such as lane-changing assist and automated parking. Mobileye said the EyeQ 6H, which will be the most advanced driver assistance chip in the family when it becomes widely available, will have three times more performance than its EyeQ 5 while only burning through 25% more power, giving it a boost for carrying out AI chores in cars.
The EyeQ 6H will begin sampling later this year and enter full production by the end of 2024.
A Long Road Ahead
At CES, the company said that it continues to land deals to supply self-driving car brains to car companies. Mobileye announced that it plans to partner with auto maker Zeekr to roll out what the companies said would be the world's first Level 4 autonomous vehicle in China by 2024.
The companies said the model would use six EyeQ 5 chips and use digital maps generated by Mobileye's Road Experience Management (REM) system, which pools data from all the cars on the road with Mobileye hardware.
Mobileye said it has also agreed to supply its road-mapping data to Ford and Volvo. The digital road maps are accessed via the cloud in real-time to provide up-to-date information that helps supplement a car's onboard sensors.
But last year, Intel said it planned to spin out a portion of its autonomous driving unit, in a bet that the car chip business will become an increasingly important market for the chip vendor. Intel, which is trying to raise funds to build chip-making fabs and close the gap with TSMC and other rivals that ended Intel's chip-making leadership, said that it plans to retain a majority stake in the company after Mobileye's IPO sometime in 2022.
Intel said spinning out the unit would give the standalone company a higher profile and the ability to lure more business. Intel said the two companies will continue to collaborate on technologies for the automotive sector.
Intel acquired Mobileye in 2017 for approximately $15.3 billion.
