Waymo Unveils Custom 5nm Compute Engine Powered by TSMC, NVIDIA, and AMD Components.

Inside Waymo Drive: Custom ASIC Architecture and 1,000 TOPS ML Compute Engine Revealed

Autonomous vehicle leader Waymo has disclosed technical details regarding the custom silicon and computing architecture powering the latest generation of its Waymo Driver autonomous system.

Transitioning away from off-the-shelf, general-purpose commercial processors, Waymo engineered custom System-on-Chip (SoC) silicon tailored specifically to meet stringent safety requirements, ultra-low latency response times, and real-time road situational processing. Designed for harsh automotive conditions, the custom platform operates reliably across extreme vibration profiles and severe temperature swings ranging from freezing sub-zero environments to blistering desert heat. The architecture also incorporates hardware-level fail-operational redundancy, ensuring secondary backup systems activate instantaneously if a primary node experiences a failure.

At the core of the computing stack is a custom Application-Specific Integrated Circuit (ASIC) manufactured on TSMC’s 5nm process node. Waymo collaborated with leading global semiconductor partners to supply key components, explicitly citing contributions from AMD, Micron, NVIDIA, Samsung, SanDisk, Socionext, and TSMC.

The primary breakthrough of the custom silicon is its integrated Machine Learning processing engine, delivering 1,000 TOPS (Trillions of Operations Per Second) of local compute power. This massive processing capacity enables the system to continuously synthesize high-bandwidth sensor streams including long-range LiDAR, radar, and HD cameras encircled around the vehicle without experiencing thermal or computational bottlenecks.

Why do off-the-shelf CPUs and GPUs struggle as the number of self-driving cars increases? While typical consumer and data center GPUs provide high floating-point computing power, they consume excessive power and generate enormous heat, requiring complex liquid cooling systems that rapidly drain electric vehicle batteries. Custom ASIC chips allow Waymo to directly assign specific neural network operations to silicon logic, achieving maximum efficiency per watt while maintaining low latency.

Traditional automotive electronics rely on "fail-safe" designs; that is, if any component fails, the system shuts down or stops safely. However, self-driving cars traveling at 65 mph on highways cannot simply shut down. Waymo's "fail-safe" silicon chip incorporates physically separated processing paths, a dual power bus, and backup sensor channels, ensuring the vehicle maintains full cognition and steering control even in the event of total main chip failure.

Instead of building completely separate proprietary hardware from scratch, which poses a significant supply chain risk, Waymo adopted a hybrid silicon strategy, designing core ML processing blocks in-house. By integrating proven IP modules and memory components from semiconductor giants such as Micron, Samsung, AMD, and NVIDIA, Waymo strikes a balance between cutting-edge performance and recognized automotive-grade manufacturing reliability.

 

Source: Waymo 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Inside Waymo Drive: Custom ASIC Architecture and 1,000 TOPS ML Compute Engine Revealed

Autonomous vehicle leader Waymo has disclosed technical details regarding the custom silicon and computing architecture powering the latest generation of its Waymo Driver autonomous system.

Transitioning away from off-the-shelf, general-purpose commercial processors, Waymo engineered custom System-on-Chip (SoC) silicon tailored specifically to meet stringent safety requirements, ultra-low latency response times, and real-time road situational processing. Designed for harsh automotive conditions, the custom platform operates reliably across extreme vibration profiles and severe temperature swings ranging from freezing sub-zero environments to blistering desert heat. The architecture also incorporates hardware-level fail-operational redundancy, ensuring secondary backup systems activate instantaneously if a primary node experiences a failure.

At the core of the computing stack is a custom Application-Specific Integrated Circuit (ASIC) manufactured on TSMC’s 5nm process node. Waymo collaborated with leading global semiconductor partners to supply key components, explicitly citing contributions from AMD, Micron, NVIDIA, Samsung, SanDisk, Socionext, and TSMC.

The primary breakthrough of the custom silicon is its integrated Machine Learning processing engine, delivering 1,000 TOPS (Trillions of Operations Per Second) of local compute power. This massive processing capacity enables the system to continuously synthesize high-bandwidth sensor streams including long-range LiDAR, radar, and HD cameras encircled around the vehicle without experiencing thermal or computational bottlenecks.

Why do off-the-shelf CPUs and GPUs struggle as the number of self-driving cars increases? While typical consumer and data center GPUs provide high floating-point computing power, they consume excessive power and generate enormous heat, requiring complex liquid cooling systems that rapidly drain electric vehicle batteries. Custom ASIC chips allow Waymo to directly assign specific neural network operations to silicon logic, achieving maximum efficiency per watt while maintaining low latency.

Traditional automotive electronics rely on "fail-safe" designs; that is, if any component fails, the system shuts down or stops safely. However, self-driving cars traveling at 65 mph on highways cannot simply shut down. Waymo's "fail-safe" silicon chip incorporates physically separated processing paths, a dual power bus, and backup sensor channels, ensuring the vehicle maintains full cognition and steering control even in the event of total main chip failure.

Instead of building completely separate proprietary hardware from scratch, which poses a significant supply chain risk, Waymo adopted a hybrid silicon strategy, designing core ML processing blocks in-house. By integrating proven IP modules and memory components from semiconductor giants such as Micron, Samsung, AMD, and NVIDIA, Waymo strikes a balance between cutting-edge performance and recognized automotive-grade manufacturing reliability.

 

Source: Waymo