Nvidia Drive Orin system-on-a-chip is built on Nvidia’s new Ampere GPU architecture
Computers have been a key component of almost every vehicle built since the 1970s when it became clear that vacuum hoses and centrifugal distributor advance mechanisms weren’t going to be sufficient to meet emission and fuel efficiency standards. Today’s vehicles typically have upwards of 100 processors ranging from tiny low-power microcontrollers to virtual supercomputers. As advanced driver assist systems (ADAS) and eventually automated driving (AD) become ubiquitous, Nvidia
After cancelling its March GPU Technology Conference (GTC) in San Jose due to the pandemic, Nvidia has been running online sessions for the past several weeks. One of the highlights of GTC has always been the epic unscripted keynote from company founder and CEO Jensen Huang that can run two and half hours or more. As with the rest of this year’s San Jose GTC, Huang’s keynote went online this time. While much of the talk this time centered around servers and gaming, ADAS and AD are also a key market for Nvidia in the coming years.
Back in December 2019, prior to CES, Nvidia announced its next-generation system-on-a-chip (SoC) for automotive applications, called Orin. Aside from the top-line specifications such as the ability to deliver up to 200 trillion operations per second (TOPS) at significantly lower power consumption, few other details were provided at the time.
At the heart of Huang’s keynote was the announcement of Nvidia’s next-generation graphics processor (GPU) architecture known as Ampere. The Nvidia Drive AGX Orin series is based on this new platform. Ampere is the successor to the current Turing architecture that powers Nvidia’s RTX 20xx series video cards. However, the current generation Nvidia Drive platforms including Xavier are based on the predecessor to Turing known as the Volta.
The Xpeng P7 is the first production application of the Nvidia Drive Xavier system-on-a-chip
The first production application of Xavier was just recently announced, the Xpeng P7 with its partially automated Xpilot 3.0 system. AD developers have also had access to the Drive Pegasus system since mid-2018 which utilizes two Xavier and two Volta GPUs. Orin is skipping over Turing and going straight to the Ampere generation.
Xavier which is rated at 30 TOPS with 30W power consumption is targeted at level 2+ systems such as GM Super Cruise or Tesla
The new Nvidia Drive Orin platform is designed to scale from entry-level driver assist systems to … [+]
New car assessment programs (NCAP) in Europe and elsewhere offer a great marketing opportunity for automakers that can achieve 5-star safety ratings. In Europe, evaluations of ADAS capabilities are becoming part of the scoring regime and automakers are trying to improve the reliability and capability of systems like automatic emergency braking and pedestrian detection.
“Our customers are requesting that we’ve been servicing level 2 platforms and above and delivering robotaxi offerings and a wide range of advanced solutions,” said Danny Shapiro, Nvidia senior director of automotive. “It’s too costly for customers to develop separate architectures for different levels of automated driving and so now we’re going to have complete entry to high-end solutions, ADAS to robotaxis.”
Nvidia’s chief rival in this sector is Intel
One of Nvidia’s selling points is a common architecture that developers can use from initial bench development to simulation to on-road testing. More importantly, as new generations of Nvidia silicon come along, they maintain compatibility so that code developed to run on Xavier or earlier platforms can move directly to Orin or whatever comes next.
Nvidia will be offering an Orin ADAS SoC that can be installed in windshield enclosures along with cameras and radar. This entry-level chip provides 10 TOPS performance at a claimed 5W of power consumption. This should be more than sufficient to power these more capable ADAS functions.
For the next-generation L2 and L3 partially automated systems, Nvidia offers the baseline Orin with up to 200 TOPS performance. That chip is projected to consume about 45W. Finally for high-end robotaxi applications, there will be a successor to the current Pegasus system. Featuring two Orin and two Ampere chips, this board is claimed to be capable of 2,000 TOPS compared to 320 TOPS for Pegasus. While it does have more than six times the processing capability of Pegasus it also uses twice the power at 800W. That’s a significant improvement in computing efficiency, but this isn’t a platform you’re going to find in any mainstream vehicles.
The Orin family of SoCs will be available for start of production from late-2022 or early 2023. Samples for development and testing will be available in 2021.
In addition to silicon, Nvidia also provides a lot of software for developers to build on. The newest addition to that software stack is Drive RC for remote control. The Drive RC components can be used to route sensor data to locations other than ADAS or AD computer for a range of applications including eventually teleoperation to support automated vehicle deployments. However, subsets can be used to build camera mirror systems that replace external mirrors on vehicles for improved aerodynamic performance. The cameras used for ADAS or partial automation can display to screens in the cockpit to give the driver a view of what’s behind.
Nvidia is already utilizing Drive RC as it resumes some of its on-road testing to enable social distancing. Like most companies developing AD systems, Nvidia has always used two safety operators in the vehicle. One keeps their hands by the wheel and watches the road, ready to take control when needed. The co-pilot watches data and tracks anomalies that need to be investigated. That co-pilot is now watching the sensor data from the vehicle in real-time using Drive RC.
By moving both down-market and at the same time raising the performance bar at the high-end for fully automated driving, Nvidia wants to make it increasingly challenging for automakers to look elsewhere for their in-vehicle computing needs. However, the raw performance numbers published by vendors like Nvidia, Intel, NXP, Qualcomm
Nvidia’s platforms have become increasingly optimized for running artificial intelligence type algorithms including deep neural networks. However, many in the AD development world are recognizing the limitations of those approaches as highlighted by former Starsky Robotics CEO CEO Stefan Seltz-Axmacher when he recently announced he was shutting down the company. Having a diversity of software including neural networks and rules-based algorithms can make for a more robust system that is also easier to validate and the raw performance of one type of chip may not necessarily provide the best overall performance in that case.
Regardless of what approach developers take, we are in for continued interesting times in the automotive electronics space.
