Analysis
Waymo has designed its own machine-learning accelerator chip, built on TSMC's 5-nanometer process and delivering more than 1,000 TOPS of AI performance, replacing the Intel FPGAs the company previously used to process sensor data, The Register reported. Waymo says it will present further technical detail on the chip at the Hot Chips conference at Stanford.
The chip's job is narrow and safety-critical: convert raw feeds from more than a dozen onboard cameras into driving responses within milliseconds, running both convolutional neural networks and transformer models that draw on more than 200 million miles of Waymo's own autonomous-driving data. Waymo says the system performs real-time temporal noise reduction to improve low-light visibility, and every vehicle carries the chip in a redundant pair with liquid cooling wired into the car's own coolant loop -- if one chip fails, the second can take over without interrupting the drive. "Within those critical milliseconds, advanced ML models build a high-fidelity understanding of the environment to evaluate the safest path forward," the company said.
Why Waymo moved off FPGAs
Waymo had previously relied on Intel field-programmable gate arrays for sensor processing, which the company found difficult to program and lacking the compute density modern transformer-based driving models require. Custom silicon solves both problems at once -- purpose-built logic runs the exact workload Waymo needs rather than a general-purpose chip repurposed for the job -- at the cost of a multi-year design and fabrication cycle that only a company with Alphabet's balance sheet behind it could absorb. Waymo isn't going fully vertical the way Tesla has with its own FSD computer: non-ML components like data orchestration, movement and logging still come from outside partners, including AMD, Micron, Samsung, SanDisk and Nvidia.
The competitive read
Tesla has run a custom silicon program for its Full Self-Driving computer for several hardware generations now, betting that a single, cheap, camera-only stack can scale to every consumer vehicle it sells. Waymo's redundant, liquid-cooled, dual-chip design is the opposite bet: safety-certified reliability first, unit cost second, aimed at a robotaxi fleet rather than a mass-market car. Both companies converging on custom silicon at the same time -- years after Tesla started -- suggests general-purpose compute has hit a ceiling for real-time autonomous driving workloads industry-wide, not just at these two companies.
The counterweight
Designing and taping out a custom ASIC is a multi-year, capital-intensive bet with real execution risk: chip revisions are slow and expensive relative to software iteration, and Waymo is entering this race years behind Tesla's silicon program. The redundant dual-chip, liquid-cooled design also implies a higher per-vehicle cost than a single-chip system, which cuts against Waymo's broader push toward cheaper, purpose-built vehicles for unit economics. Waymo has not disclosed per-chip or per-vehicle cost figures, so it's not yet possible to say whether the custom silicon actually lowers Waymo's cost structure or simply improves latency and reliability at a premium.
