Waymo Develops 5nm In-House ASIC: Why It Builds Custom Chips Despite Its Small Fleet

Waymo recently unveiled a 5-nanometer (nm) application-specific integrated circuit (ASIC) that it will install in its self-driving vehicles. It is an unusual move for a company with a far smaller fleet than Tesla to design high-performance chips in-house. The ASIC delivers more than 1,000 TOPS (tera operations per second), combining front-end processing with machine learning (ML) model performance. Waymo's overall computer system is a heterogeneous architecture that pairs this ASIC with CPUs, GPUs, and other components.

Waymo judges that custom ASIC development is economically worthwhile because the value of the autonomous driving computer in each vehicle is very high, even though its vehicle count is not as large as Tesla's. Custom chips carry high design and initial costs (NRE), but Waymo's sixth-generation system uses a vast array of sensors, including 13 cameras, 4 lidars, and 6 radars. Because the real-time computing load from this complex sensor fusion is so large, the analysis holds that custom silicon is essential to maximize power, heat, and space efficiency. Through this, Waymo aims to ensure the stable operation of its advanced autonomous driving system.

Tesla has already been mounting custom autonomous driving silicon in large volumes since the Hardware 3 (HW3) era, and it has a structure that is advantageous for spreading costs because it mass-produces vehicles itself. Tesla also advances its autonomous driving technology through a vertically integrated structure that directly controls even the software stack. This contrasts with Waymo's strategy of developing its own chips while leaving vehicle production to outside manufacturers.

Waymo and Tesla take opposing approaches to building their autonomous driving systems. Waymo aims for fully driverless operation and secures sensor redundancy, including lidar and radar, maintaining the complexity of its sensor system while trying to process it efficiently with custom computing. Tesla, by contrast, focuses on simplifying its sensor system and building a vision-centered system. Accordingly, it is difficult to judge which approach is superior based on ASIC performance alone.

The Ojai platform is based on a vehicle from Zeekr, a Chinese automaker. Through this platform, Waymo plans to bring its next-generation autonomous driving technology into real-world service.

Zeekr is confirmed to be responsible for manufacturing Waymo's vehicles in China. Meanwhile, integration of the Waymo Driver, the core of autonomous driving, takes place at the Waymo and Magna facility in Mesa, Arizona, USA. This shows Waymo's intent to focus on developing and integrating its autonomous driving system rather than getting directly involved in vehicle manufacturing.

Waymo is using the Hyundai Ioniq 5 as an autonomous driving platform in addition to the Zeekr-based robotaxi. The Waymo-equipped Ioniq 5 is produced at Hyundai Motor Group's Metaplant America in Georgia, USA. This shows that Waymo is pursuing a strategy of extending its autonomous driving system to vehicles from various manufacturers rather than being confined to a specific model. The aim is to secure system flexibility and scalability by utilizing multiple platforms.

Waymo develops its own sensors and writes its autonomous driving software independently. It also develops in-house computers containing custom ASICs optimized for that software, and integrates this entire driver system into various automotive platforms. While leaving automobile manufacturing to existing manufacturers, this strategy strengthens control in the core autonomous driving technology areas through vertical integration from sensors to AI computers. It is Waymo's distinct approach, starting from the opposite direction of Tesla, which achieves vertical integration from vehicle production to software, to secure core technological capability.

Ultimately, the key question in the Tesla vs. Waymo competition may not be whether lidar is used, whether cameras alone are used, or the TOPS performance of the ASIC. The real key lies in who first secures reliability sufficient for people not to need to drive, and how cheaply that system can be scaled up in volume. Rather than differences in technology stacks, fundamental system reliability and cost efficiency will likely be the decisive factors.