According to a post on X, Elon Musk said that Tesla has yet to unlock the full potential of AI4 – the now-renamed Hardware 4.
HW4 / AI4
AI4 currently runs Hardware 3 in emulation mode, which means it's only running about as fast as Hardware 3 can run. With just basic specs in mind, AI4’s cameras have 4-5x better overall resolution and fidelity than Hardware 3 and should be about 3-5x faster in inference capabilities. That’s a considerable increase over the baseline specs of Hardware 3.
With FSD 12.5, AI4 is running a FSD model that has a 5x increase in parameters, resulting in a smoother driving experience. Recently, Elon Musk noted that AI4 should be able to provide an additional 8x time increase in parameters. While we’re not exactly sure how the parameter metric computes vs the inference capability metric, but if we add them, it’s a 12x increase in capability.
However, Elon’s post implies that this is an 8x increase over the capacity they originally thought to have – nearly 50x faster than Hardware 3. That’s a bold claim, but could entirely be based on software optimization and not having to run Hardware 3 in emulation mode in the future, once AI4 and Hardware 3 begin to diverge.
AI4 Divergence
Divergence is still expected to be a few months away – at least not until early 2025. However, Elon confirmed that the new Supercomputer Cluster, which will run AI4 alongside Nvidia GPUs, is coming online very soon. This means that work on improving FSD with the newly unlocked potential of AI4 is going to start soon – however, it may not necessarily come to customers right away.
FSD on Hardware 3
Currently, most of Tesla’s fleet is still on Hardware 3 – the Model 3 only changed over after the refresh, and the Model Y only changed over in late 2023. Even if AI4 brings drastic improvements, Tesla will still have to focus a lot of energy on optimizing their model so that Hardware 3 vehicles – some of which are less than a year old today, can still run the latest FSD updates.
Musk recently commented on X that Tesla is still optimizing FSD to run on hardware 3. He originally estimated that it would take about 10 days to optimize the code for HW 3, however, that date has come and gone. We’re now on day 15, and HW 3 owners are getting increasingly frustrated about the delay and the worry that HW 3 is reaching its limits.
FSD v12.5.1.3 is currently the latest FSD version available to the public. However, Musk said that this version doesn’t have HW 3 support yet, and it’ll be a new unreleased version that will roll out for HW 3.
Tesla previously told its employees on the Cybertruck production line that work would be halted for three business days and that they would not be required to show up to Giga Texas for any other work. Work was halted on the Cybertruck line from December 3rd until December 6th.
Many speculated that Tesla had hit a demand mark with the Cybertruck or run into other production-related difficulties. However, pausing production for such a short period is a fairly normal business operation, as Tesla continuously retools and upgrades its production lines.
Halts for business demand tend to look more similar to Ford’s F-150 Lightning production halt - where they halted production from November 18, 2024, until January 6, 2025. Ford also announced that this production halt was due to reduced demand for its electric truck in the face of stiffer competition.
Cybertunnel
In this case, part of the reason for the halt could potentially have been due to upgrades to the underground tunnel that is used to move Cybertrucks fresh from the production line to the holding lot at Giga Texas. Tesla has been working on improving the production flow rate for its vehicle - and rather than encountering issues with production, they’ve previously mentioned they’ve had issues with getting vehicles out of their outbound parking lots and off their assembly lines - and into customers’ hands fast enough.
The Cybertunnel moves trucks from inside Giga Texas to the holding lot just outside the factory. At that holding lot, they’re either loaded into trains or trucks to be transported across North America. The tunnel passes underneath Texas Highway 130, helping to ensure that Tesla isn’t clogging up traffic crossing the highway with brand-new trucks.
More Shut Downs?
It’s likely we’ll see even more shutdowns at Giga Texas in the coming months. Tesla has shown that it intends to continue increasing production rates at Giga Texas, which means it will need time to upgrade its assembly lines and processes. In addition, the Model Y is also produced at Giga Texas—and once the Model Y Juniper arrives, there will need to be shutdowns to accommodate re-tooling.
So, no need to stress - it's just regular operations. And we get to see a cool Cyber-themed tunnel too. In the meantime, check out Joe Tegtmeyer’s drone video of Giga Texas below.
Giga Texas on this cold 6 December 2024 morning! Big news with the @boringcompany paving & tunnel covering structure on W side, new mystery construction on W side of GA, continued work on SW corner for tunnel alternate exit & production of both @cybertruck & Model Y today,… pic.twitter.com/94a894ix1I
Tesla’s FSD V13 introduced a host of new features, including the ability to Unpark itself, travel in reverse, and Park at its destination. While FSD can now navigate itself out of a parking spot or garage, there are limited options when arriving at your destination, but that’s about to change.
Garage Support
Ashok Elluswamy, Tesla’s VP of AI, confirmed on X that a future revision of FSD V13 will introduce a variety of parking options, including parking in garages, driveways, parking spots, or simply pulling over on the side of the road. Drivers will presumably be able to pick when choosing a location, or maybe the vehicle will offer suggestions based on the location.
This will be a big addition, as today, FSD doesn’t understand how to park in a garage or driveway. Some owners trick FSD into parking in their driveway by using tape or road paint to outline parking spots.
Soon, that may not be necessary, and instead, FSD will simply back itself into your garage. This feature will probably work best in larger garages—extremely tight garages may not work, given the reduced precision of vision at extremely close distances.
We hope that if Tesla introduces this feature, they’ll sell printable guides you can stick on the garage walls or floor to help ensure your vehicle is in the perfect spot every time. Many commenters mentioned that they park in urban/suburban garages with only inches to spare around their vehicles, so this will be an interesting challenge for Tesla to address.
Parking Garages
Parking garages offer their own unique problem set. For the most part, FSD can already navigate its way out of a parking garage, assuming that there aren’t any special gates or sequenced/signaled signage.
It is also able to find a parking spot in multi-level garages and navigate its way out. More support for multi-level parking garages is coming, as FSD seems to stick to the garage walls until it can find the exit itself. It seems that Tesla’s data gathering and vehicle-to-fleet communication could aid Tesla in figuring out parking garages in terms of knowing the entrances and exits.
FSD 13.2 enters a parking garage. I then park it on the second floor and it escapes the parking garage. I then park it on the snow covered roof and... pic.twitter.com/GxfsldfkUz
In the Upcoming Improvements section of FSD V13, Tesla mentions that more destination options will be supported, including pulling over (at the side of the road) and parking in a spot, driveway, or garage.
These other destination options are going to be important soon - because one of the biggest FSD features we’re all still waiting for is Banish AutoPark. Once your vehicle can pull over at the entrance of a location and then go park itself, we’re going to be super close to Unsupervised FSD becoming a reality.