The story making the rounds this month is that FSD v14 Lite is tied to rising HW3 computer failures. Owners reporting overheat and fault errors, mostly on hot summer drives. My reaction was not surprise. I drive a 2021 Model S Long Range with HW3 and paid FSD, so this is not an academic topic for me, and I’ve been posting pieces of this analysis in owner threads for a couple of years now. This post is the assembled version.

The three computers

First, some context, because “the computer in a Tesla” is actually three computers, and people mix them up constantly:

Infotainment. Intel or AMD depending on the year. Runs the screens and the audio. This is the one that reboots when you hold the scroll wheels.

Gateway. A custom low-power ARM processor that is on all the time. It owns the in-vehicle networks (LIN and CAN) and brokers commands to the other vehicle control modules. You never think about it, which is the point.

The AP computer. Two custom inferencing processors and all the cameras. This is what takes over when you turn on cruise or FSD, and it’s the thing everyone means when they say HW2, HW3, HW4, or HW5.

Since the Model 3 launched, all three have been packaged together as a single compute package sharing an on-package databus. Same idea across the fleet, VERY different form factors across the generations. That last part matters a lot in a minute.

The wattage tells the story

In my day job I size AI infrastructure, so when I look at the AP computer generations I don’t see car parts, I see accelerators:

Generation	Package power	Inferencing is said to be roughly a…
HW3	200-300W	Nvidia T4
HW4	400-450W (and ~75% larger by footprint)	RTX 4070
AI5 / HW5	projected 600-700W	RTX 5090 / Blackwell class

Those are datacenter numbers riding around behind your dash! And they explain a bunch of otherwise-random engineering decisions, like why Tesla moved the cars from 12v to 16v around the HW4 transition (more watts through the same gauge wire), and why the Cybertruck went all the way to 48v. The power delivery, the cooling loop, and the harness are all sized to the package. Remember that when we get to the retrofit question.

Why the failures aren’t a mystery

The HW3 board has a bunch of redundancy built into it. Two inferencing nodes, with the workload designed to run at around 50% capacity so that either side can take over if the other faults. Full failover, by design. And even in normal operation, Memory, Storage, and Turbo processor errors occur routinely on these boards. They’re aging, and they were built with headroom precisely so that aging wouldn’t matter.

Then v14 Lite arrived. Even the distilled model needs so much compute and memory that HW3 runs the package well past the 50% design point. Maybe not a flat 100%, but judging by the failures, close. The failover headroom is mostly gone, and the computer is dumping close to double the planned heat into a thermal loop that is already taxed on a hot summer day. So you get overheat and fault errors, concentrated in exactly the conditions owners are reporting.

In datacenter terms: they’re running the N+1 cluster at N, at the top of its thermal envelope, in August. Anyone who has watched a passively-cooled T4 throttle in a chassis with lazy airflow knows how this movie ends! (My mining-era bench taught me more about thermal throttling than any datasheet ever did.) This is also why unsupervised driving is simply never coming to HW3 (Tesla finally admitted as much this spring). The redundancy budget and the compute budget are the same budget, and v14 has already spent most of it.

The retrofit math

Tesla has said, in various wordings over the years, that cars with paid FSD will get an upgrade path. Okay. What would that actually take? Walk the stack with me:

Power. Going from a 200-300W package to a 400W+ or 700W package is not a board swap. You’re into the Power Conversion System and the vehicle controllers (VCBatt, VCFront), plus jumper and adapter harnesses to bridge the old connectors to the new package.

Permutations. How many versions of HW3 shipped across Model 3, Y, S, and X over the years? The number is pretty nuts. Every permutation is its own retrofit engineering problem, with its own kit and service procedure. On my own car’s platform alone there’s the 8GB computer, a rare 16GB revision of the same board, and then the “Gaming” E through G revisions that widened the eCall connector to sneak Ethernet pins in. I ordered the update jumper harness and tried it on the older revision myself. It doesn’t work.

Packaging. This is the part people hand-wave. On the Model S, both generations of car computer live in the same place (passenger footwell, behind the footrest panel), and that’s about where the compatibility ends. Go pull up the service manual: HW3 secures with two bolts and one nut, HW4 with two bolts and two nuts, on a different pattern, and the stud that locates an HW3 sits right about where it would puncture an HW4. The coolant plumbing moved too. HW3’s coldplate ports sit close together near the middle of the unit; HW4 spread them out to the sides, so a swap means re-plumbing the thermal loop, on a procedure that already requires a full coolant drain (with a five-hour clock on the drain routine) and a vacuum refill.

The connectors moved around too: HW3’s Autopilot board has ten connectors plus a separate gaming-computer connector, HW4’s has eight and no gaming connector at all. Nobody has published measured dimensions, but the teardown folks who have had both units on a bench call HW4 a totally different form factor, thinner overall but with a bigger board footprint. I’ve been telling people who ask about computer swaps on HW3 Palladiums the same thing for over a year now: not going to happen, the sheetmetal and plumbing isn’t compatible. There isn’t really a way to make it fit.

Cameras. HW3 cameras have the resolution of a potato next to the 5MP HW4 set, and the newer cameras run different voltage and cabling. And Tesla now concedes the new computer needs the new cameras (that admission finally came this April), so the “computer swap” is really a computer, power, harness, and camera project. On a car that was never wired for any of it.

Feature parity. And there’s a newer wrinkle: even inside HW4, the computer quietly got worse at the end. Every Model S since early 2023 is an HW4 car, but the ones built after roughly June 2025 lost the discrete AMD GPU on the infotainment side and had storage halved from 256GB to 128GB, which converged the Model S computer on the Cybertruck’s configuration (the truck never had the GPU, and never got Steam). New S and X deliveries had already lost Steam back in May 2024, via a quiet message to customers waiting on their cars, while everyone already delivered kept it. So the last and “best” Model S computers ever built do less than the 2023 units do, and with S and X production ending this past quarter, nobody is putting it back. My September 2021 build has the 8GB infotainment computer, so I never had Steam to lose (Sad Panda), but I followed the 16GB retrofit threads for a year. And there’s precedent for upgrades that subtract: ask the HW2.5 crowd from the last computer-swap era, whose companion MCU upgrade cost them their broadcast radio, plus $500 more to get FM and XM back. AM never returned.

Which brings us to the silicon roadmap, because the retrofit target keeps moving. AI5 was supposed to be in cars in the second half of 2025 (that was the promise at the June 2024 shareholder meeting). Instead, the design was declared finished in July 2025, followed by Saturday design reviews all fall, followed by “almost done” in January 2026, with volume production pushed to mid-2027 and tape-out finally happening this April. Read between those lines however you like. My read is the part didn’t meet the need and they started over, and given that a chip takes about two years from design start to volume silicon (with roughly a year of that after tape-out), the dates line up.

Meanwhile the stopgaps arrived: AI4.5, a three-chip unit that quietly started shipping in Model Ys in January, and the AI4+ announced in April with double the memory per chip (the official framing is that AI5 goes to Optimus and the datacenters first, which is its own tell). So what does that mean for HW3 cars? A repackage was going to be required for any retrofit anyway, because nothing newer fits the HW3 envelope. There’s a VERY good chance that AI4+ repackaging work is what eventually goes into HW3 cars. Tesla has even floated dedicated retrofit micro-factories, because the service centers would choke on the volume. That’s the space I’m watching.

Running down the clock

Meanwhile, notice what Tesla is actually doing: free FSD transfers to newer cars, offered again and again. Every owner who transfers retires a retrofit liability at zero hardware cost. Elon’s favorite saying is that the best part is no part, and the cheapest retrofit is the one you never have to ship. So my read, and I’ve been saying this in the forums for a while, is that they’re running down the clock on those of us holding paid FSD on HW3 cars.

Now think about the conundrum for the folks deepest in this hole. Say your car started life as HW2.5. You bought FSD. You already went through one computer swap to get HW3, and if you were on the old infotainment, the MCU upgrade that swap wanted is the one that took your radio. Musk has promised the next upgrade free to FSD buyers and called it “painful and difficult.” Painful and difficult toward WHAT, exactly? Do you get AI4? AI4+? AI5? The most anyone will commit to is AI4-something plus the cameras, and the answer decides whether the thing you paid for, twice now in hardware terms, ever actually shows up.

If you’re shopping used right now, the advice writes itself: buy the newest car you can afford, and if the budget lands you on an HW3 car, make sure FSD is an included package so you’re in line for whatever hardware update eventually materializes. If it never does, well, you didn’t pay extra for it.

As for my car: the early Palladium Model S is the most upgradable Tesla ever made, and I’ve already done a bunch of the incremental updates the newer cars shipped with (that instinct has its own origin story, and the tow hitch quest is ongoing). If an AI4+ retrofit program ever opens, I’ll be first in line, and you can bet I’ll document every connector and part number here! Until then the car stays supervised, and on hot days I know exactly what that computer is going through. 🙂

Thanks for reading!