Tesla AI Platform Features 'Two Parallel Brains'
Tesla's latest AI4 platform reportedly introduces an architecture with "two parallel brains" to power both its Full Self-Driving (FSD) system and its Optimus robotics program. This design approach reflects a trend toward modular and redundant real-time computing for safety-critical autonomous systems. In a related update, Tesla is also rolling out voice prompts that allow drivers to issue verbal commands for FSD actions like parking.
- The AI4 platform, also known as Hardware 4.0 (HW4), is built on Samsung's 7nm process and features 20 ARM Cortex-A72 CPU cores, an increase from 12 in the previous generation. This design choice prioritizes reliability and cost-effectiveness over utilizing the most advanced manufacturing processes. - The "two parallel brains" concept refers to a dual-redundancy architecture where two separate computing modules run in parallel to continuously verify each other's computations, ensuring a fail-safe for critical FSD functions. This approach is a common strategy in safety-critical autonomous systems to eliminate single points of failure. - This unified hardware approach allows advancements in AI and neural network processing to be shared across both Tesla's vehicles and its Optimus humanoid robot. The Optimus robot leverages the same foundational FSD computer and neural networks for its perception and decision-making. - Tesla's journey to in-house chip design began after its partnership with Mobileye, which supplied the EyeQ3 chip for Hardware 1, ended in 2016. Tesla then briefly used NVIDIA's Drive PX2 platform for Hardware 2 before introducing its own custom-designed FSD chip with Hardware 3 in 2019. - A key competitor to Tesla's AI4 is NVIDIA's Drive Thor platform, which is built on a more advanced TSMC 4N process and boasts a theoretical performance of up to 2,000 TFLOPS. While AI4 uses older ARM Cortex-A72 CPU cores, Thor utilizes modern, server-grade ARM Neoverse CPUs. - The upcoming FSD v13 software is a major architectural overhaul designed to take advantage of the AI4 hardware, featuring enhancements like 3x scaling of the model size and context length, and a 5x increase in training compute power. - Voice commands in Teslas use a natural language processor, allowing drivers to control features like climate, navigation, and media without using rigid phrases. Commands such as "I'm cold" or "Speed up the wipers" are understood by the system. - The Optimus robot, currently in its Gen 2 iteration, stands at 168cm, weighs 57kg, and is deployed in Tesla factories for tasks like sorting battery cells. It utilizes the same vision-based AI as Tesla vehicles, relying on cameras to build real-time 3D maps without LiDAR.
