Check out more coverage of GTC Fall 2022.
NVIDIA introduced a new automotive-grade SoC to serve as the central computing hub of self-driving cars.
The Santa Clara, California-based company said the Drive Thor pumps out up to 2,000 TOPS of performance using the company’s new FP8 data format, a major generational leap over its current Orin SoC series.
The chip features NVIDIA’s new multi-instance “Hopper” GPU to handle the machine-learning workloads at the heart of autonomous cars, plus it also brings its latest “Lovelace” GPU architecture into the fold. In addition, Thor adds a high-performance Arm-based “Grace” CPU. With 77 billion transistors, the supercomputer-class SoC is designed to unify the clusters of computer systems that control modern cars into a single platform.
Danny Shapiro, head of NVIDIA’s automotive business, said Thor keeps it a step ahead of automakers that need high-performance hardware to enable more automated safety and self-driving functions in their cars.
“Autonomous cars are one of the complex computing challenges of our time,” said Shapiro. “With safety being paramount, nobody’s ready to release these vehicles into the wild until there’s more computing.”
NVIDIA is wrestling with Qualcomm (with its Snapdragon Ride suite) and Intel (with its Mobileye EyeQ SoCs) to convince automakers to use its self-driving chips. It has rolled out several generations of its Drive SoCs that can manage the safety-critical workloads underpinning automated and assisted driving. It also offers silicon to control dashboard displays, digital instrument clusters, camera mirrors, and infotainment systems.
While it usually takes a bevy of different chips to control all of these systems, NVIDIA said Thor has enough processing power that automakers can effectively consolidate many of their functions into a single chip.
NVIDIA signaled that Thor is still a work in progress for now. It will enter volume production in 2024 and hit the road in 2025 vehicle models.
A New Architecture
As software-powered features become a bigger focus for automakers, so has the hardware under the hood.
Today, more than 100 electronic control units (ECUs) can be distributed throughout a modern vehicle. Each module generally only has enough computing power to handle a single task, such as a parking-assist system. But as the car’s complexity gets out of control, automakers are moving to “domain-based” architectures that combine many of these single-use modules into “domain controllers” that can be upgraded over time.
High-performance chips equipped to each control unit are designed to run several different functions safely all at once instead of separate microcontrollers (MCUs). The systems run in separate software containers.
Other companies are upgrading to “zonal-based” architectures, where a central onboard computer is linked to the sensors and other systems through “gateways” that communicate data around the car over Ethernet.
NVIDIA said Thor is suited for centralized architectures where many cameras, radar, and sensors, and even displays, are connected directly to the platform without the use of intermediary chips to pre-process data.
“We can do sensor fusion directly instead of relying on any intermediary,” said Shapiro. But he added that if its customers prefer to attach processors directly to the sensors, Thor will give them the flexibility to do so.
Performance = Safety
Putting everything on a single chip requires a huge amount of computing power that Thor promises to offer.
Thor uses an automotive-grade version of the high-performance “Poseidon” CPU cores being developed by Arm for data centers, giving it access to one of the most advanced central processing cores on the market. The chip supplies up to 8X the performance of NVIDIA’s Orin SoC to process the vast amounts of data from cameras, radar, and other sensors on autonomous cars and then plot out a safe route on the road ahead.
Last year, NVIDIA introduced a new automotive-grade SoC called “Atlan” that was expected to provide 1,000 TOPS of performance on INT8 when it launched in 2024. However, the company said it has canceled the Atlan chip in favor of Thor.
While NVIDIA is not announcing many of the details about its architecture, Thor will probably have many of the same building blocks as Atlan. But it also leverages many of the latest capabilities from the company’s GPUs.
Thor includes a new inference engine specifically designed for “transformers.” This is a new type of machine-learning technology that is rapidly replacing convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for AI. Transformers process videos as a single frame, giving Thor the power to handle more data over time.
The transformer engine—a new component inside the tensor cores at the heart of NVIDIA’s server GPUs—can improve the performance of machine learning models based on transformers by up to 10X, according to the company.
But when you integrate all of these different features into a single architecture, you need secure hardware isolation to keep safety-critical workloads and non-safety-related ones from interfering with each other.
To safely consolidate these disparate systems, the Thor SoC is also capable of domain-based computing. Thus, it can partition itself so that safety-critical workloads can execute without interruptions or delays.
Furthermore, the technology enables the chip to run several operating systems at the same time. For instance, the car’s core operating system could run on Linux, while the digital dashboard runs on QNX or even Android.
This also opens the door for customers to funnel all of Thor’s performance into the autonomous driving pipeline or use a portion (maybe 1,000 TOPS) to run the dashboard display and use the remainder for ADAS.
NVIDIA said Thor and the AGX system based on it are both designed to meet the ASIL-D standard for functional safety under the ISO 26262 standard. The software stack has both ISO 26262 and ASPICE compliance.
The hardware and software are also designed to meet automotive security standards, including ISO 21434.
While Thor will likely cost more than Atlan, NVIDIA said its customers should come out ahead when it comes to system-level cost savings since the car’s electrical and electronics architecture is simplified with a single SoC.
“You can imagine tremendous savings in terms of cost, in terms of reduced cabling, in terms of reduced weight, in terms of reduced energy consumption overall,” said Shapiro in a briefing with reporters. “Then there is the ease of allowing a single software update to provide new features across these different ECUs.”
On top of that, the lack of cabling reduces the need for connectors that can be safety hazards if they are shaken loose.
One SoC, Many Configurations
NVIDIA plans to roll out different configurations of the Thor SoC, ranging from a single-chip solution to a “superchip” that connects two Thor SoCs using its NVLink-C2C interconnect technology to run a single unified operating system. The company said it would give automakers the headroom to continuously upgrade their cars over time with new services and even additional safety features via over-the-air updates à la Tesla.
“We will have a range of different options available, so customers will be able to choose the right level of performance for their needs,” said Shapiro. “But it’s up to them to decide what the right configuration will be for them based on their needs and the sensors on their cars.” Thor is designed to scale up from advanced driver-assistance systems (ADAS) such as lane-changing assist to fully autonomous driving, according to the company.
The power efficiency of chips is a major requirement for electric cars, where they must compete for limited battery life. NVIDIA said Thor is 3X more efficient than Orin, without sharing the specific numbers on power.
Thor is paired with the same Drive software development kit (SDK) as Orin, which when coupled with its scalable architecture, allows companies to port their past software development to the new platform.
Check out more coverage of GTC Fall 2022.