What semiconductors lurk beneath?
Now that Tesla Inc. is sustaining production of its latest electric car, the Model 3, the decidedly unconventional automaker is seeking to upgrade key functions like its autopilot by moving beyond graphics processors to an internally developed “AI chip.” The approach reflects the massive computing power required for future releases of the Tesla autopilot. Analysts say that it also represents an enormous financial and technical risk for the debt-burdened car manufacturer.
Tesla’s plan is to swap out older GPUs from current supplier Nvidia used as the main processor to run its autopilot algorithms. They would be replaced by an internally developed ASIC designed as a neural-network accelerator. Observers said that Tesla is betting that a dedicated AI chip would prove more efficient than general-purpose GPUs for specific software applications and inference.
The so-called AI chip was disclosed by Tesla founder Elon Musk during an Aug. 1 conference call with analysts to discuss the carmaker’s quarterly results. “We’ve been like semi-stealth mode basically for the last two to three years on this, but I think it’s probably time to let the cat out of the bag because the cat’s going to come out of the bag anyway,” said Musk.
The disclosure was followed by a series of performance claims by Musk about what he billed as nothing less than “the world’s most advanced computer designed specifically for autonomous operation.” Market watchers note that Musk and his chip design team provided few details about the three-year-old internal development effort other than to say that their AI chip could outperform an unspecified Nvidia GPU by a factor of 10.
“This hardware replacement does not automatically increase the performance of autopilot without accompanying software upgrades,” noted auto analyst Phil Magney of VSI Labs. “It gives Tesla more headroom to push out future software updates that can take advantage of the faster computer.”
Magney and other Tesla watchers nevertheless doubt the performance comparisons because the carmaker is not using Nvidia’ latest-generation processor. For example, the main board in Tesla’s Autopilot 2.0 uses only one Nvidia Parker SoC and one Pascal GPU rather than Nvidia’s latest-generation Xavier-based SoCs or Volta GPU cores, noted Magney.
Musk further asserted that “using a GPU, fundamentally it’s an emulation mode, and then you also get choked on the bus. So the transfer between the GPU and the CPU ends up being one of the constraints of the system.”
Auto analysts note that the current Tesla architecture does not include Nvidia’s NVLink Fabric, the company’s new high-speed GPU interconnect. Nvidia claims that NVLink allows data to flow between the processors up to 12 times faster.
Besides Nivida, other component suppliers for the Tesla Autopilot 2.0 include Samsung (DRAM), Marvell (Ethernet switches), and Infineon (microcontroller), according to VSI Labs (see chart on page 3).
Another teardown posted by the website Ingineerix probed one of the three body controllers used in the Model 3. The fairly straightforward board included many discrete components along with Tesla ASICs and STMicroelectronics’ Power Architecture chip.
“Tesla has kind of just implemented all this stuff discretely,” the website determined, calling it an “awesome design [that] completely replaces the fuse box and power distribution in the front of the car.”
While the autopilot remains a major focus of Tesla trackers trying to determine its next move, there are several other intriguing upgrades in the Model 3. Some are driven by internal chip development while others represent a new take on standard features, including a patented air-flow system. The novel HVAC system allowed Tesla engineers to remove mechanical louvers and vents in order to narrow the dash, thereby maximizing space in the Model 3 cabin.
The patented cabin ventilation system uses “two intersecting planes of air,” according to a Tesla engineer quoted in this video. In what amounts to a marketing gimmick, it also includes high-efficiency particulate air (HEPA) filters.
Meanwhile, the Model 3’s widely admired battery management system includes internally developed ASICs and other Tesla-designed IP. After checking the voltage on the Model 3’s 80-kWh battery pack, a bench engineer pronounced the design “absolutely gorgeous, just stunning.”
While teardown jockeys are generally impressed with the guts of the Model 3, the automated vehicle stack in general and the autopilot in particular continue to draw the most interest and debate, largely due to Musk’s audacious performance claims.
Steve Jobs envy
Mike Demler, senior analyst with The Linley Group, countered that Tesla’s risky strategy of developing an AI chip from the ground up “makes absolutely no sense financially.” He and others also doubt whether an expensive in-house hardware development effort is justified given the availability of other powerful processors from Mobileye and Nvidia. Demler suspects that Tesla silicon engineers prefer a “closed system,” a kind of pride of ownership.
“Elon Musk has Steve Jobs envy,” he quipped.
The focus on developing an ASIC for neural-network acceleration is misguided, insists Demler, explaining that real innovation in automotive electronics actually resides within software. “Sensors, sensor fusion, and intelligence are really what’s important,” he said the day after Musk rattled financial markets with a proposal to take Tesla private.
“We’ll have to wait and see when they go to silicon” for the Tesla AI chip, he added, but “it’s not likely they are doing anything groundbreaking.” One reason is that established players like Nvidia and Mobileye are more than meeting demand in the automotive electronics sector for ever-greater processing horsepower with platforms, including Nvidia’s single-chip Xavier “brain of the self-driving car” unveiled in May and its Pegasus AI computer designed for self-driving taxis and other autonomous vehicles.
Observers agree that safe autopilots will require huge amounts of processing power both for perception and cognition; that is, using algorithms to understand the rules of — as well as the forks in — the road.
“We need to stop thinking about autonomous vehicles and start thinking about autonomous modes” along the line of use cases like adaptive cruise control, said Demler.
The pros and cons of Tesla’s ASIC approach continue to be debated, especially given the development cost of what Demler called “reinventing the custom wheel.” While ASICs essentially freeze a chip design at fabrication, auto analyst Phil Magney notes that the architecture remains more efficient than GPUs. One reason is because ASICs are “optimized for a particular application” — in the case of the Model 3, AI-based inference models, he added.
“Nvidia SoCs and companion logic serve many applications very well that require massively parallel instructions,” said Magney.
Demler does give Tesla credit for driving the transition to electric cars, especially its battery research. “That’s their key innovation,” he said. According to another Ingineerix teardown of the power conversion system, the 12-V assembly located under the rear seat serves as the AC-to-DC on-board charger and DC converter. The entire assembly is mounted on one large PC board. The board also contains perhaps the only fuses used in the Model 3.
These and other power management components have enabled Tesla to offer a long-endurance battery that provides more than 300 miles of driving range, a key metric for moving electric cars into the mainstream.
In the following pages, we will share some of the key components used in Tesla 3 uncovered by several teardowns.
The battery management system (BMS) board
Power conversion system
Front body controller
— George Leopold is the former executive editor of EE Times and the author of Calculated Risk: The Supersonic Life and Times of Gus Grissom (Purdue University Press, 2016).
New products & solutions, whitepaper downloads, reference designs, videos
Register, join the conference, and visit the booths for a chance to win great prizes.