Genma_Jp
@nymbusjp · Mar 28, 2026
If you want to know who has a chance to compete with FSD, apart from the nasty quantization tail that can only be solved with Hardware In the Loop for training, you should also look at the compute.
Elon mentioned that you need one H100 to simulate one HD camera at real-time
Elon Musk@elonmusk· Mar 19, 2026@jhong @c_valenzuelab We’ve been able to generate physics-accurate, real-time video for self-driving training & testing at @Tesla_AI for a long time.
The compute required for this (roughly one H100 per HD camera) is still far too expensive for consumer use, but probably becomes affordable in 2 to 3
Elon Musk
@elonmusk
Simulation is used to reproduce extremely rare or dangerous scenarios where we don’t have enough data, despite millions of cars on the road.
For example: you’re driving down a narrow two-lane road when a pedestrian trips and falls into the road. The right decision is to slow down as much as possible and choose a glancing impact into the oncoming car, rather than run over and kill the pedestrian.
This is the classic Trolley Problem, but with a vast number of variations.
11:08 AM · March 28, 2026 · 280.1K views
397
419
6K