Waymo developed its own chip for robotaxis and released its list of suppliers.

      Waymo has revealed details about the contents of its robotaxis' trunk. A blog post published on Thursday outlined the structure of its onboard computer, introduced a custom 5-nanometer chip it designed itself, and listed the seven suppliers it collaborates with to create the system. The company describes this as the first glimpse into its trunk, which is significant for a company that typically shares very little about its hardware.

      What the chip actually does is an area where the coverage and the company differ, so it’s important to be clear. The chip is an application-specific integrated circuit (ASIC) created using a 5nm process. Waymo states that it processes “the massive influx of raw data before it reaches our core ML brain.” It retrieves information from raw lidar, radar, and camera feeds, including temporal denoising for improved perception in low light, and then sends that data to a separate inference engine that operates the sensor-fusion models. Essentially, it functions as a front-end processor, preparing what the driving system observes rather than determining the vehicle's actions.

      This distinction is critical for the performance figures cited. Waymo claims that “these ASICs alone deliver over 1,000 TOPS of ML performance dedicated to front-end processing and ML models.” Bloomberg noted that, based on this metric, Waymo's performance would be on par with Nvidia's latest autonomous-driving systems. This comparison involves one component performing a singular task versus entire driving platforms, but it is not a claim made by Waymo.

      Waymo has not eliminated Nvidia from its partnerships, nor has it moved away from AMD; both are included in the supplier list along with Micron, Samsung, Sandisk, Socionext, and TSMC. The company describes a “balanced, heterogeneous system” that integrates its own machine-learning silicon with what it deems the top CPUs, GPUs, and accelerators available. The custom chip provides additional capability rather than replacing the commercial parts surrounding it.

      Waymo also indicated that the ASIC is “just one of several exciting custom components we’re developing.” More proprietary silicon is forthcoming, although the company hasn't disclosed specifics.

      Waymo's system is built around three main requirements: responsive, ruggedized, and redundant. The first requirement, latency, has led the company to coining the term “pixels-to-actuation” to describe the delay from when a photon hits a sensor to when the car reacts. In the past eight years, their raw computing power has increased twentyfold to reduce this latency.

      The second requirement is physical durability. The computer is installed in a car, not a server rack, meaning it must withstand vibrations, shocks, and extreme temperature fluctuations from Midwest winters to Phoenix summers. Waymo connects it to the vehicle’s liquid cooling system, ensuring it operates silently while allowing for cargo space.

      The third requirement is particularly noteworthy. According to the post, “Our compute is designed like two independent engines.” They typically function together handling full parallel workloads, but if one encounters a malfunction, the other assumes control. There is no human fallback, so the redundancy exists in the hardware itself.

      Regarding scale, Waymo highlights that it has accumulated more than 200 million miles of fully autonomous experience informing this design. Currently, it operates about 4,000 vehicles across more than ten cities, reporting around 500,000 paid trips per week, as noted by The Verge. All these numbers come directly from Waymo. The company also asserts that its technology reduces accidents and injuries compared to human drivers, based on its own safety data, which has not been independently verified.

      The system can process data from 13 high-resolution cameras simultaneously, which corresponds to the Ojai, a purpose-built four-seater also equipped with four lidar sensors. Zeekr, a subsidiary of the Chinese automaker Geely, manufactures the Ojai. TNW reported in May that the vehicle is more cost-effective to produce, more durable, and produced in China. Bloomberg has indicated that the custom chip is now being installed in Ojai vehicles, although specifics on quantity and timelines have not been disclosed.

      Political considerations are relevant here, as Washington has been tightening regulations concerning Chinese vehicle hardware throughout the year. One supplier, identified by the Pentagon as a military entity, still employs Chinese lidar in American robotaxis. Waymo’s collaboration with Geely—where Waymo designs the intelligence while Geely manufactures the vehicle—serves as a response to this challenge, though neither company has explicitly framed it this way.

      Just a day before the chip announcement, Waymo made the Ojai available to all riders in Los Angeles, Phoenix, and San Francisco, as reported by Engadget, after previously limiting it to select invitations.

      Waymo is not acting in isolation; its parent company has been working on silicon design for years to manage AI infrastructure costs. Additionally, this past Wednesday, Marvell granted Google a $12.2 billion share option under a custom-chip agreement.

      One notable comparison involves Waymo's