📰The Verge•Freshcollected in 1m
Waymo Reveals Its Robotaxi Compute Brain
💡Waymo exposes the specialized compute stack powering commercial driverless vehicles.
⚡ 30-Second TL;DR
What Changed
The robotaxi computer can perform up to one quadrillion operations per second.
Why It Matters
The disclosure offers developers and researchers a rare look at the compute infrastructure required for commercial autonomous driving. It may also intensify competition around specialized automotive AI hardware and system design.
What To Do Next
Review Waymo’s hardware disclosure and compare its compute architecture with your autonomous-driving or robotics workload requirements.
Who should care:Developers & AI Engineers
Key Points
- •The robotaxi computer can perform up to one quadrillion operations per second.
- •Waymo revealed details about its chip architecture, processors, and internal hardware components.
- •The company also published a list of suppliers supporting its driverless fleet.
🧠 Deep Insight
Background and context from public sources — not the original article. 27 sources cited.
🔑 Enhanced Key Takeaways
- •Waymo's robotaxi compute system integrates server-grade CPUs and GPUs to process the massive sensor data streams.
- •The volume of Waymo's compute engine has been successfully reduced in its latest vehicles, providing more usable trunk space for passengers.
- •The fifth-generation Waymo Driver, introduced in March 2020, was designed as a platform-agnostic system capable of powering various vehicle types and use cases, including Waymo One for ride-hailing and Waymo Via for logistics.
- •Waymo's sixth-generation Driver, launched in February 2026, significantly reduces hardware costs by approximately 40% by pushing more processing complexity into custom silicon chips and relying less on multiple discrete hardware components.
- •The 6th-generation system utilizes fewer cameras (13 compared to 29 in the 5th-generation) due to the adoption of higher-resolution 17-megapixel imagers and custom-designed chips, while maintaining or improving performance.
📊 Competitor Analysis▸ Show
Competitor Compute Hardware Comparison
| Feature/Company | Waymo (6th-Gen Driver) | Tesla (Hardware 4) | Mobileye (EyeQ™6H) | NVIDIA (DRIVE Thor) |
|---|---|---|---|---|
| Compute Power (TOPS/FLOPS) | Up to 1 quadrillion operations per second (1000 TOPS) / 200 TFLOPS (edge) | 3-8x more powerful than HW3 (HW3 was 144 TOPS) | 12 TOPS (EyeQ5) / EyeQ6H is Mobileye's most advanced SoC | 2000 TOPS (announced for next-gen AVs) |
| Chip Architecture | Custom silicon chips, server-grade CPUs and GPUs | Custom "FSD Computer 2" SoC (7nm process) / Custom "FSD Chip" (14nm FinFET CMOS) | EyeQ™ System-on-Chip (SoC) | DRIVE AGX platform |
| Sensor Approach | Multi-modal fusion: 13 cameras (17MP), 4 lidar, 6 radar, external audio receivers | "Pure vision" (cameras only) | Multi-sensor fusion (cameras, radars) / True Redundancy (independent camera and radar-lidar subsystems) | Supports various sensor configurations (platform) |
| Memory | Not explicitly detailed, but handles 25TB/day sensor data | 16GB RAM, 256GB storage (HW4) / 32GB RAM per SoC (64GB total for AI4+) | Not explicitly detailed | Not explicitly detailed |
| Cooling | Liquid cooling (speculated for Google's TPUs, which Waymo may use) | Not explicitly detailed | Not explicitly detailed | Liquid cooling (for high-performance AV platforms) |
| Power Consumption | Requires significant edge processing | Not explicitly detailed | Not explicitly detailed | Not explicitly detailed |
🛠️ Technical Deep Dive
- Waymo's onboard computer combines server-grade CPUs and GPUs, forming the 'brain' of the Waymo Driver.
- The system processes approximately 25 terabytes of sensor data per vehicle daily, requiring an edge processing capability equivalent to 200 TFLOPS.
- Waymo's perception stack is capable of processing 20GB/s of data with an end-to-end latency of 3 milliseconds.
- The 5th-generation Waymo Driver utilizes a comprehensive sensor suite including 29 cameras, lidar, and radar systems.
- The 6th-generation Waymo Driver features a streamlined sensor configuration with 13 cameras (using 17-megapixel imagers), 4 lidar units, 6 radar units, and external audio receivers.
- Waymo's custom-designed chips and optical designs, with core components developed and built in California, are central to the 6th-generation system's efficiency and performance.
- The Waymo Foundation Model employs a "Think Fast and Think Slow" (System 1 and System 2) architecture, comprising a Sensor Fusion Encoder for rapid reactions and a Driving VLM (Vision-Language Model) for complex semantic reasoning.
- The Sensor Fusion Encoder merges camera, lidar, and radar inputs over time to produce objects, semantic attributes, and embeddings for fast driving decisions.
- The Driving VLM component uses rich camera data and is fine-tuned on Waymo's driving data to handle rare situations requiring background world knowledge.
- Waymo's training infrastructure leverages 50,000 TPUs to process 14 million hours of driving data.
🔮 Future ImplicationsAI analysis grounded in cited sources
Waymo's continued investment in custom silicon and cost optimization will accelerate its geographical expansion and market penetration.
Lower hardware costs, exemplified by the 40% reduction in the 6th-generation system, make deployment more economically viable in new and diverse markets.
The 'Think Fast and Think Slow' architecture, integrating a Driving VLM, will enable Waymo to handle increasingly complex and rare driving scenarios with enhanced safety.
By combining rapid sensor fusion with deep semantic reasoning, the system can better interpret nuanced situations and react more appropriately to unforeseen events.
Waymo's strategy of designing both hardware and software in-house will continue to provide a significant competitive advantage in system optimization and safety.
Tight integration allows for co-optimization of components, leading to more efficient processing and robust performance across diverse driving conditions.
⏳ Timeline
2009-01
Google Self-Driving Car Project begins, the precursor to Waymo.
2016-12
The Google Self-Driving Car Project rebrands as Waymo, an Alphabet company.
2017-02
Waymo introduces its custom-built, fully-integrated hardware suite for its Chrysler Pacifica Hybrid minivans, moving away from reliance on off-the-shelf components.
2017-09
Waymo announces a partnership with Intel to power its automated driving stack, utilizing Intel's CPUs and FPGAs.
2020-03
Waymo unveils its fifth-generation Waymo Driver, featuring a redesigned sensor suite with imaging radar and a reduced compute engine volume for more trunk space.
2026-02
Waymo begins fully autonomous operations with its 6th-generation Driver, featuring custom silicon chips, a 40% reduction in hardware costs, and fewer, higher-resolution cameras.
📎 Sources (27)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- builtin.com
- waymo.com
- sfdesignweek.org
- eetimes.com
- withwaymo.com
- waymo.com
- waymo.com
- cleantechnica.com
- businessmodelcanvastemplate.com
- waymo.com
- bytebytego.com
- introl.com
- wikipedia.org
- teslahubs.com
- eeworld.com.cn
- mobileye.com
- autopilotreview.com
- wikipedia.org
- marketsandmarkets.com
- reddit.com
- waymo.com
- waymo.com
- wikipedia.org
- medium.com
- forbes.com
- medium.com
- waymo.com
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Verge ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.