Thor SoC increases processing power for future AVs
Check out more coverage from GTC Fall 2022.
NVIDIA has introduced a new automotive-grade SoC to serve as the central computing hub of self-driving cars.
The Santa Clara, Calif.-based company says the Drive Thor pumps out up to 2,000 TOPS of performance using the company’s new FP8 data format, a major generational leap from its current Orin SoC series.
The chip incorporates NVIDIA’s new multi-instance “Hopper” GPU to handle machine learning workloads at the heart of self-driving cars, and also incorporates its latest “Lovelace” GPU architecture. Additionally, Thor adds a high-performance Arm-based “Grace” processor. 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 is staying ahead of automakers who need high-performance hardware to enable more automated safety and self-driving features in their cars. .
“Self-driving cars are one of the complex computing challenges of our time,” Shapiro said. “With safety being paramount, no one is ready to release these vehicles into the wild until there is more reckoning.”
NVIDIA is battling with Qualcomm (with its Snapdragon Ride suite) and Intel (with its Mobileye EyeQ SoCs) to convince automakers to use its standalone chips. It has deployed several generations of its Drive SoCs capable of handling the safety-critical workloads that underpin 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 slew of different chips to control all of these systems, NVIDIA said Thor has enough processing power that automakers can effectively pack many of their functions into a single chip.
NVIDIA reported that Thor is still a work in progress at this time. It will enter volume production in 2024 and hit the road in 2025 vehicle models.
A new structure
As software features become a priority for automakers, so does the hardware under the hood.
Today, more than 100 electronic control units (ECU) can be distributed in a modern vehicle. Each module usually only has enough computing power to handle a single task, such as a parking assistance system. But as the complexity of the car spirals out of control, automakers are turning to “domain-based” architectures that combine many of these single-purpose modules into “domain controllers” that can be upgraded to the over time.
The high-performance chips equipped with each control unit are designed to safely perform several different functions simultaneously instead of separate microcontrollers (MCUs). The systems run in separate software containers.
Other companies are moving to “zone-based” architectures, where a central on-board computer is linked to sensors and other systems through “gateways” that communicate data around the car via Ethernet.
NVIDIA said Thor is suited for centralized architectures where many cameras, radars and sensors, and even displays, are plugged directly into the platform without using intermediate chips to pre-process data.
“We can do sensor fusion directly instead of relying on an intermediary,” Shapiro said. But he added that if his customers prefer to attach processors directly to sensors, Thor will give them the flexibility to do so.
Performance = Security
Putting everything on a single chip requires massive computing power which Thor promises to deliver.
Thor uses an automotive-grade version of the high-performance “Poseidon” processor cores developed by Arm for data centers, giving it access to one of the most advanced central processing cores on the market. The chip provides up to 8 times the performance of NVIDIA’s Orin SoC to process the large amounts of data from cameras, radars and other sensors on self-driving cars, then plot a safe route on the road ahead.
Last year, NVIDIA introduced a new automotive-grade SoC called “Atlan” that was expected to deliver 1,000 TOPS of performance on INT8 when it launches in 2024. However, the company said it canceled the Atlan chip in favor of Thor.
Although NVIDIA isn’t announcing many details about its architecture, Thor will likely have many of the same core elements as Atlan. But it also leverages many of the company’s latest GPU features.
Thor includes a new inference engine specifically designed for “transformers”. It is a new type of machine learning technology that is rapidly replacing convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for AI. Transformers treat videos as a single image, 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 transformer-based machine learning models by up to 10 times, according to the company.
But when you integrate all these different features into a single architecture, you need secure hardware isolation to prevent security-critical and non-security workloads from interfering with each other. .
To securely consolidate these disparate systems, the Thor SoC is also capable of utilizing domain-based computing. Thus, it can partition itself so that security-critical workloads can run without interruption or delay.
Additionally, the technology allows the chip to run multiple operating systems at the same time. For example, the car’s central operating system could run on Linux, while the digital dashboard runs on QNX or even Android.
It also opens the door for customers to funnel all of Thor’s performance into the self-driving pipeline or use some (perhaps 1,000 TOPS) to run the dashboard display and use the rest 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 ISO 26262. The software stack is ISO 26262 and ASPICE compliant.
Hardware and software are also designed to meet automotive safety standards, including ISO 21434.
While Thor will likely cost more than Atlan, NVIDIA said its customers should lead in system-level cost savings because the car’s electrical and electronic architecture is simplified with a single SoC. .
“You can imagine huge savings in terms of cost, in terms of reduced wiring, in terms of reduced weight, in terms of reduced power consumption overall,” Shapiro said during a briefing with reporters. . “Then there is the ease of allowing a single software update to deliver new functionality across these different ECUs.”
In addition to this, the lack of wiring reduces the need for connectors which can be a safety hazard if jerked around.
One SoC, many configurations
NVIDIA plans to deploy 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 will give automakers the leeway to continually improve their cars over time with new services and even additional safety features via Tesla-style over-the-air updates.
“We will have a range of different options available, so customers can select the right level of performance for their needs,” Shapiro said. “But it’s up to them to decide what the right setup will be for them based on their needs and the sensors in their cars.” Thor is designed to transition from advanced driver assistance systems (ADAS) such as lane change assist to fully autonomous driving, according to the company.
Chip energy efficiency is a major requirement for electric cars, where they must compete for limited battery life. NVIDIA stated that Thor is 3x more efficient than Orin, without sharing specific power figures.
Thor comes with the same Drive SDK as Orin, which, together with its scalable architecture, allows companies to transfer their past software developments to the new platform.
Check out more coverage from GTC Fall 2022.