Tesla HW3 FSD lite version months away

🔑 Enhanced Key Takeaways

• •The 'FSD Lite' initiative is a response to mounting class-action litigation and regulatory pressure regarding Tesla's failure to deliver 'full' autonomy on the aging Hardware 3 (HW3) platform.
• •Technical constraints of the HW3 chip, specifically its limited memory and compute capacity compared to the newer HW4 and AI5 platforms, prevent the deployment of the full end-to-end neural network stack used in newer vehicles.
• •Tesla is utilizing a 'distilled' version of its latest FSD v13+ model, which involves pruning neural network weights to fit within the HW3 memory footprint while maintaining supervised functionality.

🛠️ Technical Deep Dive

• •HW3 (Tesla FSD Computer) utilizes dual custom-designed SoCs, each containing a Neural Processing Unit (NPU) capable of 72 TOPS, totaling 144 TOPS.
• •The 'Lite' version implements a model distillation process where a larger 'teacher' model (trained on HW4/AI5 compute) transfers knowledge to a smaller 'student' model optimized for the HW3 NPU instruction set.
• •Memory bottleneck: HW3 is limited by its LPDDR4 memory bandwidth, which restricts the complexity of the transformer-based vision models currently used in Tesla's end-to-end architecture.
• •Software stack: The update requires a specific kernel-level optimization to manage the NPU's context switching more efficiently, as the newer FSD models require higher concurrency than the original HW3 design anticipated.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline