Tesla has published a new patent that could keep its older Hardware 3 (HW3) Full Self-Driving computers useful for years longer than many drivers feared. The filing, US20260017503A1, describes a way to run higher-precision AI models on chips that were originally built for low-bit integer math.
HW3 was launched in 2019 with custom neural network accelerators tuned for 8‑bit integer operations. Each FSD computer carries two neural processors, with MAC arrays that can process thousands of operations in parallel and deliver roughly 144 TOPS, which was strong for convolutional neural networks at the time.
Newer FSD stacks, like the v13 and v14 “world model” style systems, lean on larger transformer architectures and more detailed occupancy networks that benefit from 16‑bit or 32‑bit precision, plus a lot more memory bandwidth. As models grew, memory demand rose too. FSD v13’s core driving logic needs several gigabytes more than v12, which strains HW3’s resources.
Owners of HW3 cars worried that this hardware ceiling would leave them behind newer vehicles built on HW4 and future AI5 platforms. Some investors and lawyers have already flagged the gap between early “FSD capable” marketing and the practical limits of older computers.
How Bit‑Augmented Arithmetic works
The new patent centers on what Tesla calls “Bit‑Augmented Arithmetic Convolution.” In plain terms, it breaks high‑precision numbers into smaller low‑bit pieces so existing 8‑bit MAC units can process them step by step, with results recombined later to recover a higher‑precision outcome.
The method splits a 16‑bit value into two 8‑bit parts: a most significant byte that holds coarse detail and a least significant byte that carries fine detail. HW3 then runs several 8‑bit multiply accumulate passes, similar to the FOIL pattern in algebra, and uses bit shifting and addition to stitch these partial results back into something close to a native 16‑bit operation.
Another Tesla patent, US20260017019A1, focuses on high‑precision rotary positional encoding on low‑bit hardware, which is important for transformer attention. The system converts angular data into a logarithmic domain so narrow data buses can transport high‑accuracy values, then uses Taylor series in a high‑precision stage to recover accurate trigonometric values for positional encoding.
Smart use of existing MAC hardware
A third filing, US20260017051A1, addresses how to pack more information into limited registers and data paths using the MAC units themselves. Instead of adding new packing circuitry, the patent describes repurposing the same multipliers already present in the AI core so they do double duty: math plus data packing.
By multiplying an 8‑bit value by a weight such as 256 (2⁸), the result is pushed into the upper half of a wider accumulator, leaving space to insert another value in later cycles. A small safety gap between packed values protects sign and carry bits, which helps keep numbers stable when they are unpacked and reused by neural networks.
Commentary on the filing notes that this approach lets Tesla treat the MAC blocks as both compute engines and high‑speed data movers, cutting extra wiring density and easing thermal hot spots on the chip. Engineers see this as a way to keep HW3 running heavier math without a new layout.
Impact on FSD versions for HW3
These patents give Tesla a way to train one high‑precision FSD model in the data center and then deploy adjusted versions on several hardware tiers, from HW3 to HW4 and the coming AI5. Higher‑end chips can run the full model, while HW3 receives lighter variants that still benefit from newer architectures but fit into its compute and memory budget.
Tesla already took a smaller step in this direction with FSD (Supervised) v12.6 on HW3, which brought in parts of newer logic using heavy optimization. Reports now point to a planned “v14 Lite” for HW3, expected around 2026, that would lean directly on these mixed‑precision and packing tricks to keep older cars closer to the newer stack.
Still, the company has told owners that HW3 will lag behind HW4 on feature rollout and capability, even with these tricks in place. More complex perception and planning features will likely remain exclusive to newer hardware that has more TOPS, wider memory buses, and higher‑resolution camera feeds.
Limits, latency and camera constraints
The patent authors agree that this approach comes with trade‑offs. Running what amounts to several 8‑bit passes to emulate a 16‑bit operation costs extra time and energy per inference step, so latency and power draw both rise compared with native 16‑bit or 32‑bit hardware.
Camera hardware is another hard limit. HW3 cars use roughly 1.2‑megapixel cameras, whereas HW4 moved to about 5‑megapixel sensors with more detail and better low‑light performance. Software can stretch what the existing sensors deliver, but cannot invent detail that is not captured in the first place, which may matter for long‑range detection and small distant objects.
Mixed‑precision and quantization techniques are common tools in AI deployment, yet this patent stack adapts them tightly to Tesla’s own FSD chips and driving workload. For drivers, that translates into extended support but not full equality between hardware generations.
Retrofit promises and AI5 context
All of this sits beside a separate issue: long‑term hardware upgrades. In early 2025, Elon Musk said Tesla would replace the computers in HW3 cars at no extra cost for customers who bought the FSD package outright, though not for subscribers. The company has not locked in a public timetable and still needs to solve practical questions around power, cooling, and physical integration.
HW4 uses more power, has different connectors and a different form factor, so a drop‑in swap into older vehicles is not straightforward. Some analysts expect a retrofit path based on the newer AI5 platform once that hardware is in volume production and Tesla has clearer data on the compute budget needed for unsupervised autonomy.
In the meantime, HW3 remains on the road in an estimated several million cars built from 2019 onward, and many owners still rely on subscription or purchased FSD features. For that group, these patents act as a bridge: they hold performance closer to newer stacks and may buy time until any large‑scale retrofit program becomes real policy rather than a promise.
You may also like to read:
- Tesla to launch lighter FSD V14 for hardware 3 by Mid-2026 »
- HW3 Tesla reacts quicker than humans in box truck close call »
- Tesla Confirms Reverse Feature for HW3 Vehicles »
We’ve launched an official Tesla Owners Online Forum. Join early and claim your "Founding Fuel" badge.
