Tesla Moves to Create Its Own Chip, but Is It All Much Ado about Nothing?

Subscribe To Download This Insight

2Q 2019 | IN-5488

On April 22—Tesla Autonomy Investor Day—Tesla made a series of announcements regarding its plans for vehicle autonomy and smart mobility. This included plans for full-self driving features on upcoming electric-vehicles, a fleet of robotaxis and the announcement for a new chip, labelled the Tesla Full Self-Driving (FSD) chip. The Tesla announcement to design its own chip for vehicle autonomy (and manufactured by Samsung) was an interesting move from an Original Equipment Manufacturer (OEM) that currently uses a modified Nvidia Drive PX for vehicle autonomy. Given the leading approach taken by Tesla toward novel technologies such as electric vehicles, software in-vehicle, over-the-air, and other features, could Tesla’s move toward a custom chip spark an industrywide change?

Registered users can unlock up to five pieces of premium content each month.

Log in or register to unlock this Insight.

 

Tesla Announces New Chips

NEWS


On April 22—Tesla Autonomy Investor Day—Tesla made a series of announcements regarding its plans for vehicle autonomy and smart mobility. This included plans for full-self driving features on upcoming electric-vehicles, a fleet of robotaxis and the announcement for a new chip, labelled the Tesla Full Self-Driving (FSD) chip.

The Tesla announcement to design its own chip for vehicle autonomy (and manufactured by Samsung) was an interesting move from an Original Equipment Manufacturer (OEM) that currently uses a modified Nvidia Drive PX for vehicle autonomy. Given the leading approach taken by Tesla toward novel technologies such as electric vehicles, software in-vehicle, over-the-air, and other features, could Tesla’s move toward a custom chip spark an industrywide change?

A Battle for Efficiency

IMPACT


Currently the automotive industry is largely split on how much and what level of computational performance measured by Tera Operations per Second (TOPs) autonomous vehicle applications will require. Autonomous driving systems vary greatly in its setup. For example, sensor suite makeup or the method of redundancy employed, and therefore inherently the amount of computation power required, will vary greatly system by system.

Nvidia and Tesla have long believed that autonomous driving systems require large amounts of power while other market incumbents such as Mobileye, AIMotive, ARM, and others have continually promoted a more “lightweight” philosophy that centers on maximizing efficiency rather than on raw computational power. Power efficiency is seen as important for two primary reasons:

  • System Capital Costs: Excessive power will lead to excessive costs for chipset and System on Chip (SoC).
  • Overall Energy Consumption: Higher power chips will require higher levels of cooling, ultimately drawing power from the electric battery. Considering future autonomous vehicles are almost exclusively likely to be battery electric vehicles, this impacts the range of the vehicle.

Power efficiency seems to be a message that now resonates with market incumbents involved in the autonomous space. The choice for Tesla therefore to adopt a more “efficient solution” that emphasizes power requirement over total computation power could therefore be a significant development. Tesla’s chip delivers 144 TOPs, drawing around 72 watts (W) of power, while Nvidia’s current top end solution, the Pegasus AGX, delivers almost twice as much computational power at 320 TOPs. However, the Pegasus AGX draws around 500W of power. Tesla’s new solution therefore requires 0.5W per TOP while NVidia’s solution requires 1.6W per TOP.

Indeed, as autonomous technology continues to develop and the idea of having driverless vehicles on the road in commercial applications becomes a reality, an increasing number of OEMs are likely to become more concerned with the efficiency of the system. So the question is: what does this mean for the market moving forward? Will every OEM spend hundreds of millions of dollars designing its own Application-Specific Integrated Circuit (ASIC), as Tesla have done, or are there other strategies moving forward?

Will Other OEMs Follow Tesla's Lead?

RECOMMENDATIONS


Nvidia so far has developed a huge head start on the market for computing for autonomous applications. The company has announced numerous high-profile deals with Volvo, Veoneer, Toyota, Audi, Continental, and others, including Tesla; the deal incorporated Nvidia chips in Model S and X vehicles, powering Tesla’s semiautonomous system, the Tesla AutoPilot.

Tesla’s choice therefore to move away from the Nvidia Xavier architecture toward its own ASICs is an interesting move. However, perhaps more relevant than the technical discussion over chip performance and efficiency is the discussion around what Tesla plans to do with its software and hardware. ABI Research finds it highly unlikely that Tesla is likely to sell its chips or autonomous driving software anytime now or in the future to third parties such as other OEMs or tier one suppliers. This would remove the scope for differentiation for both Tesla and OEMs via software; it is unlikely that either party would want this.

ABI Research also finds it highly unlikely that other OEMs will follow Tesla’s lead and move to develop their own chips in-house. From a business model point of view, Tesla operates like no other OEM in that it focuses on in-house development. Traditional OEMs on the other hand work with tier one developers who in turn work with tier two developers and software developers. Although OEMs are increasingly moving toward in-house development, particularly for autonomous driving software, none have currently made announcements regarding developments of their own ASIC, and ABI Research finds this unlikely to change moving forward. Nvidia, with its more general purpose platform, therefore still represents the best solution for OEMs regarding vehicle autonomy.

The key takeaway from the Tesla announcement is the importance of system efficiency. Even if Tesla is capable of and/or is willing to move to in-house development of chips to achieve this efficiency, OEMs will soon (if not already) be using efficiency as the key differentiator among different hardware solutions. ASICs are clearly more efficient than graphics processing units or field-programmable gate arrays, as the chip is essentially designed for a specific task and is therefore optimized for that specific task. However, on the flip side, the solution is not as general purpose as, say, the Nvidia Xavier platform would be. This means that Nvidia can provide a nonspecific platform that meets all customer requirements rather than a chip that is highly optimized for a specific task or process. Where ASICs may gain ground in the market is where the chip is developed in conjunction with software developers. One example of such a company is AImotive. It originally developed algorithms on the Nvidia platform but now offers its own semiconductor Intellectual Property (IP), aiWare, in conjunction with its own full stack software offering, AiDrive.

Finally, although many commentators have been quick to point out the substantial different in watts per TOP between the Drive Pegasus and the Tesla FSD chip, they should not forget the current evolution of Nvidia products. The Nvidia Xavier SoC delivers performance comparable to the previous generation, the Nvidia PX2, but at a lower level of energy consumption. Given this trend, Nvidia Orin could be set up to deliver Nvidia Pegasus–level performance but at a lower energy consumption. Further, Nvidia’s unified architecture means that customers can take networks and software developed for previous platforms and export them to the latest architecture—for example, from AGX Pegasus to Orin.

Services