Tesla pushed FSD V12.5 with update 2024.20.10 to additional vehicles last night. For the first time, FSD v12.5 also went out to additional Tesla models, however, it’s still limited to hardware 4 vehicles.
Support for FSD V12.5
FSD V12.5 is only available to vehicles with hardware 4 vehicles, or as Tesla is now calling the hardware, AI4. Before yesterday the update was limited to just Model Y vehicles on HW4, but Tesla quickly expanded it to all vehicles besides the Cybertruck and Semi, which has yet to see any Autopilot features.
Musk commented on X that he wanted to focus on improving the software on specific hardware before expanding it to additional hardware and vehicles.
FSD v12.5 has a 5X increase in operating parameters. This is a large increase and Musk said that it currently won’t be able to run on hardware 3 until the code is optimized.
We are focusing on just Model Y with HW4 for the initial release. Make sure that works well, then broaden.
— Elon Musk (@elonmusk) July 23, 2024
This has the 5X increase in parameters.
HW3 would run the same parameter count, but requires extra work to optimize the code.
Ashok Elluswamy, Head of AP at Tesla, also mentioned that the focus for V12.5 was on both safety and smoothness for this release. His personal test – not spilling an open cup of coffee while on FSD. On V12.5, Ashok made it 30 minutes without spilling coffee. This is good news, as the smoother and safer FSD gets, the more natural it feels for the driver and occupants in other vehicles.
HW4/AI4
Following all this up is a recent post on X from Elon mentioning that Hardware 4 will now be renamed to AI 4. Tesla is already training the HW4 computers in a loop with Nvidia GPUs, at a 1:2 ratio. This is surprising news, as Musk previously mentioned that training for Hardware 4/AI4 would only begin after the new Supercomputer cluster was finished.
Update 2024.20.10
Percent of Fleet
Installs on Oct 17
It's good to see that training is already underway for AI4, but this also means we’re hitting the limits of Hardware 3 pretty fast. It has been mentioned that there will be a drastic hardware divergence in the future between Hardware 3 and AI4, but we weren’t expecting that until later this year.
This means that updates may get better, as Tesla’s AI4 is approximately 4-5x better in terms of resolution and fidelity than hardware 3, and hardware 4.0 is about 3-5x faster and more powerful in inference capabilities.
AI5
With AI5 scheduled 18 months away, including a 10x inferencing improvement, we’re wondering what the future of Hardware 3 and AI4 will look like within the next 3 to 5 years. Cars are long-time purchases for many people, and we’re hoping that Tesla continues to do their best work to optimize code wherever possible to continue supporting Hardware 3.
Tesla has specifically come out and said it doesn’t plan to support a retrofit from Hardware 3 to Hardware 4, likely due to the differing locations of electrical and coolant connections, as well as the form factor of the compute units themselves.
