Tesla patents 'clever math trick' for HW3, but nothing points to promised FSD
11 hours ago
- #Tesla
- #Hardware
- #Self-Driving
- Tesla has published a new patent titled 'Bit-Augmented Arithmetic Convolution' to enhance performance of its HW3 self-driving computers.
- The patent describes a method to emulate higher-precision AI models on HW3's 8-bit hardware by splitting 16-bit numbers into two 8-bit pieces.
- Despite the clever engineering, the patent does not address HW3's critical limitations, such as insufficient memory (8 GB RAM) and lower-resolution cameras (1.2 MP vs. HW4's 5 MP).
- Tesla's FSD v13 requires approximately 7.5 GB of RAM per node, leaving HW3 vehicles stuck on older software versions (v12.6) while HW4 runs v14.
- The patent acknowledges potential latency issues, which could degrade user experience and autonomous driving performance.
- Tesla's 'V14 Lite' update for HW3, expected around Q2 2026, will be a watered-down version of the already-supervised FSD v14, not fulfilling the promise of unsupervised self-driving.
- Elon Musk admitted in January 2025 that HW3 cannot support unsupervised self-driving, but Tesla has no public plan for hardware retrofits or refunds.
- Tesla's website has changed its language from 'all cars have self-driving hardware' to 'designed for autonomy,' acknowledging the original promise was incorrect.
- The patents represent a temporary solution to extend HW3's usability but do not resolve the fundamental hardware limitations or deliver on Tesla's original FSD promises.