What Tesla's Full Self-Driving Computer 3 Actually Is
The Full Self-Driving Computer 3 (internally designated HW3) is Tesla's third-generation onboard neural-network processing unit, introduced in March 2019 as a retrofit-capable upgrade to the earlier HW2.5 platform. It powers every Tesla Model 3, Model Y, Model S, and Model X built between April 2019 and roughly mid-2023, when Tesla began transitioning new production to the Hardware 4 (HW4/AI3) architecture. As of early 2026, millions of Tesla vehicles on U.S. roads still rely on HW3 as their primary autonomy compute backbone.
Understanding what HW3 can and cannot do is critical for two buyer segments: used-EV shoppers evaluating FSD-equipped Teslas, and current owners wondering whether their HW3 vehicle will ever achieve the hands-off "robotaxi" experience Tesla has long promised. The short answer, grounded in SAE automation definitions and NHTSA regulatory data, is that HW3 remains a Level 2 driver-assistance platform — and Tesla's pivot to a camera-only sensor suite imposes hard ceiling on its autonomous capability relative to lidar-equipped robotaxis like Waymo.
Tesla FSD (Supervised) — whether running on HW3 or HW4 — is classified as SAE Level 2 driver assistance. The human driver must remain engaged, eyes on the road, and legally liable at all times. This is fundamentally different from SAE Level 4 robotaxi services like Waymo One, which operate without any human driver in designated geofenced zones. Tesla's announced "Robotaxi" service has not received Level 4 regulatory approval in any U.S. jurisdiction as of 2026. (SAE J3016 Standard)
HW3 Technical Specifications and Compute Architecture
Tesla designed HW3 around a custom 14nm system-on-chip (SoC) — the Tesla FSD Chip, fabricated by Samsung — running in a dual-redundant configuration. Here is the precise hardware profile:
| Specification | HW3 (Computer 3) | HW4 (AI3/AI4) |
|---|---|---|
| SoC Process Node | 14nm (Samsung) | 7nm (Samsung/TSMC) |
| Neural Network Processors | 2x NPU (dual redundancy) | 2x NPU (higher-density) |
| Peak INT8 Inference | ~144 TOPS (72 TOPS per chip) | ~300-500 TOPS (est.) |
| Camera Inputs Supported | 8 cameras (1.2 MP each) | 8-11 cameras (up to 5 MP) |
| Radar Support | Yes (originally); removed in 2021+ | No native radar (4D radar optional in some markets) |
| Ultrasonic Sensors | 12x USS (removed Oct 2022+) | None (Tesla Vision only) |
| Power Draw | ~72W typical | ~50-65W (improved efficiency) |
| Production Vehicles | Apr 2019 – mid-2023 | Mid-2023 – present |
| Retrofit Availability | N/A (came installed) | Not offered as HW3 retrofit |
The dual-chip redundancy was a key HW3 innovation: both NPUs process the same camera data simultaneously, and if their outputs disagree, the system defaults to the more conservative driving decision. This was Tesla's first hardware platform designed from the ground up for neural-network inference rather than repurposed mobile-SoC architecture.
Sensor Stack: The Camera-Only Bet and Its Consequences
Tesla's approach to autonomous perception diverges sharply from every other robotaxi operator. The HW3 sensor suite, particularly in vehicles produced after Tesla's "Tesla Vision" transition (radar removed mid-2021, ultrasonic sensors removed late 2022), relies exclusively on eight surround cameras:
- 3 front cameras (main, wide-angle, narrow telephoto) — mounted behind windshield
- 2 front-side repeater cameras — front fenders
- 2 rear-side cameras — B-pillars
- 1 rear-view camera — above license plate
- Resolution: 1.2 megapixels per camera (HW3-era); HW4 upgraded to 5 MP
- No lidar, no radar, no ultrasonic proximity sensors
This camera-only architecture is the single most consequential technical limitation of HW3 for autonomous operation. Without lidar's direct depth measurement or radar's all-weather object detection, HW3 must infer 3D spatial geometry entirely from 2D image data using neural networks — what Tesla calls its "occupancy network" approach. In controlled testing and real-world deployment, this creates measurable performance gaps:
- Low-visibility conditions: Heavy rain, fog, direct sun glare, and nighttime unlit roads degrade camera performance. NHTSA incident reports show a disproportionate share of Autopilot/FSD-involved crashes occur in low-light or adverse-weather conditions. (NHTSA Standing General Order)
- Close-range detection: Without ultrasonic sensors (removed in late 2022), parking and low-speed maneuvering rely entirely on camera inference, resulting in documented phantom-braking and object-miss events in FSD builds throughout 2023-2025.
- Depth estimation at speed: At highway velocities, camera-only systems have less margin for error in judging the speed and distance of distant objects compared to lidar's direct time-of-flight measurement.
FSD Supervised on HW3: Real-World Capability in 2026
As of early 2026, Tesla's FSD (Supervised) software — typically running v13.x builds on HW3 vehicles — delivers the following capability profile on public roads:
| Feature | Status on HW3 | Driver Required? |
|---|---|---|
| Highway lane-keeping and auto lane-change | Functional | Yes — hands on or eyes monitored |
| On/off-ramp navigation | Functional | Yes — confirmation often required |
| City street traffic-light/stop-sign response | Functional | Yes — intervention frequent at complex intersections |
| Unprotected left turns | Functional but inconsistent | Yes — high intervention rate |
| Construction zone navigation | Limited | Yes — frequent takeover needed |
| Parking lot traversal | Basic | Yes — USS removal degraded close-range |
| Hands-off / eyes-off driving | Not approved or capable | N/A — SAE Level 2 only |
Tesla's camera-based system requires the driver to remain attentive through a cabin-camera monitoring system that tracks eye gaze. If the driver looks away for an extended period, FSD issues escalating warnings and will eventually disengage, bringing the vehicle to a stop with hazard lights activated. This is functionally equivalent to GM's Super Cruise or Ford's BlueCruise in terms of SAE classification, though Tesla's system operates on a far broader set of roads (any mapped road vs. pre-mapped divided highways).
HW3 vs. Level 4 Robotaxis: A Direct Comparison
The keyword "full self-driving" in Tesla's product name creates persistent confusion. Here is how Tesla FSD on HW3 compares to actual Level 4 robotaxi services currently operating in the United States:
| Attribute | Tesla FSD (Supervised) on HW3 | Waymo One (6th-Gen Driver) |
|---|---|---|
| SAE Level | Level 2 (driver-assist) | Level 4 (no driver required) |
| Sensor Suite | 8 cameras only (1.2 MP) | LiDAR + cameras + radar + external audio |
| Geofencing | None (any road) | Mapped service areas only |
| Driver Required | Yes — eyes on, legally liable | No — fully driverless |
| Weather Limits | Degrades in rain/fog/glare | Handles moderate rain; suspends in severe weather |
| Current U.S. Cities | Nationwide (as driver-assist) | Phoenix, San Francisco, Los Angeles, Austin (select zones) |
| Revenue Model | $8,000 software purchase or $99/mo subscription | Per-ride pricing (comparable to Uber/Lyft) |
| Regulatory Approval | Approved as ADAS (Level 2) | CPUC + state DMV permits for driverless commercial service |
| Safety Oversight | NHTSA defect investigation (ongoing) | CPUC/NHTSA reporting requirements |
The critical distinction: Waymo vehicles carry no steering wheel for the passenger and no human driver, and they are commercially permitted to charge for rides. Tesla FSD on HW3, regardless of software version, requires a licensed driver in the seat who can be cited for inattention or held liable in a crash.
HW3 and the Tesla Robotaxi Promise: Where Things Stand
Tesla unveiled its dedicated "Cybercab" robotaxi prototype in October 2024, showing a two-seat vehicle with no steering wheel or pedals, designed for a future Level 4 ride-hailing network. However, several realities constrain HW3 owners from participating in this vision:
- Cybercab uses next-generation hardware. The production Cybercab (expected 2026-2027 volume production) will use Tesla's latest AI compute platform, not HW3. Tesla has not announced any pathway for HW3-equipped consumer vehicles to join a driverless fleet.
- Regulatory approval does not exist. No U.S. state has approved Tesla for driverless commercial ride-hailing as of early 2026. The California Public Utilities Commission (CPUC), which governs the largest robotaxi market, has issued driverless permits only to Waymo and, previously, Cruise (suspended after the October 2023 San Francisco pedestrian-drag incident).
- HW3 compute ceiling. At ~144 TOPS, HW3's inference throughput is roughly one-quarter to one-third of HW4's capability. As Tesla's neural networks grow more complex — particularly the end-to-end models introduced in FSD v12+ — HW3 is increasingly the bottleneck. Tesla has not committed to a HW3-to-HW4 retrofit program, despite Elon Musk's earlier statements suggesting one would be provided if HW3 proved insufficient for full autonomy.
- Insurance and liability framework. No U.S. insurer currently offers a policy that covers a Tesla operating without a supervising driver. Level 4 commercial services like Waymo carry dedicated commercial liability policies; consumer Teslas require an attentive driver as a condition of coverage.
- Not self-driving. FSD (Supervised) is SAE Level 2. You are the driver. You are liable.
- Phantom braking. NHTSA has received thousands of complaints regarding unexpected deceleration on Autopilot/FSD, particularly on HW3 vehicles post-USS removal.
- Complex intersections. Unprotected left turns, multi-lane roundabouts, and temporary construction zones require frequent driver intervention.
- Adverse weather. Heavy rain, snow, fog, and direct sun glare reduce camera-only perception reliability.
- No guaranteed upgrade path. Tesla has not confirmed HW3 will receive all future FSD features or that a hardware retrofit will be offered.
Safety Data: What the Numbers Show
Evaluating the safety of Tesla's HW3-powered FSD requires separating Tesla's own claims from independent regulatory data:
- Tesla's claim (Q4 2025 safety report): 1 crash per ~4.5 million miles driven on Autopilot, vs. U.S. average of ~1 crash per 670,000 miles (NHTSA national average). Caveat: Tesla's data includes highway miles where Autopilot is predominantly used — a lower-risk environment than the national average mix of urban/rural/highway driving.
- NHTSA Standing General Order data: Between July 2021 and the most recent reporting period, Tesla vehicles operating on Autopilot or FSD were involved in over 1,000 reported crashes, including 25+ fatalities. NHTSA opened a defect investigation (EA24-002) into Autopilot/FSD driver-monitoring adequacy. (NHTSA SGO Reports)
- IIHS research: Studies consistently show that partial-automation systems (Level 2) can induce driver complacency, with drivers of Tesla, GM Super Cruise, and similar systems spending more time eyes-off-road than with systems requiring continuous steering-torque input. (IIHS ADAS Research)
- Intervention rate (independent testing): Consumer Reports and other independent evaluators have measured FSD (Supervised) intervention rates of approximately 1 driver takeover per 10-30 miles in mixed city/highway driving on HW3 vehicles — significantly higher than on HW4 vehicles running the same software version.
For context, Waymo's Level 4 fleet reported approximately 5.7 million fully driverless miles through mid-2024, with a crash rate roughly 85% lower than human benchmarks in comparable urban environments (per Waymo's published safety data, which is subject to ongoing independent verification). The comparison is not apples-to-apples — Waymo operates in geofenced, pre-mapped urban zones with extensive sensor suites, while Tesla FSD operates everywhere with cameras only — but it illustrates the gap between Level 2 driver-assist and Level 4 driverless performance.
Should HW3 Owners Expect a Retrofit or Trade Up?
This is the practical question for the roughly 2+ million Tesla owners with HW3 vehicles. The economics and technical realities break down as follows:
If you already own a HW3 Tesla and have FSD purchased: You will continue to receive FSD (Supervised) software updates, and the system will incrementally improve within HW3's compute constraints. However, Tesla has prioritized HW4/AI4 development for its most advanced neural-network architectures. Expect HW3 FSD to remain a capable Level 2 highway and city assist — not a pathway to unsupervised driving.
If you're shopping for a used Tesla with FSD: A HW3 vehicle with FSD purchased (the $8,000 one-time buy or active subscription) carries a price premium in the used market, but this premium has been declining as buyers recognize the HW3/HW4 performance gap. As of early 2026, used HW3 Model 3/Y vehicles with FSD typically command $1,500-$3,000 more than identical HW3 vehicles without FSD — well below the original $8,000 purchase price, reflecting market skepticism about HW3's autonomous-driving ceiling.
If you want the closest Tesla experience to "self-driving": A 2024+ HW4 vehicle running the latest FSD (Supervised) build offers meaningfully better perception (5 MP cameras, higher inference throughput, improved cabin monitoring). But it remains Level 2. For true driverless rides today, your only option is summoning a Waymo in Phoenix, San Francisco, Los Angeles, or Austin.
Frequently Asked Questions
Is Tesla's Full Self-Driving Computer 3 capable of Level 4 autonomy?
No. HW3 is a Level 2 driver-assistance platform. It requires an attentive human driver at all times. Tesla has not received regulatory approval for Level 4 driverless operation in any jurisdiction, and the HW3 compute architecture (~144 TOPS, 1.2 MP cameras) is below the threshold most experts consider necessary for unsupervised driving.
Will Tesla offer a HW3 to HW4 hardware upgrade?
As of early 2026, Tesla has not announced or offered a HW3-to-HW4 retrofit program. When Tesla transitioned from HW2.5 to HW3, it offered a free retrofit to FSD purchasers. No equivalent program exists for HW3 owners. If Tesla eventually launches a consumer robotaxi network, HW4 (or newer) hardware would almost certainly be required.
Can I use FSD on HW3 without paying the subscription?
Tesla's basic Autopilot (Traffic-Aware Cruise Control and Autosteer) is included on all HW3 vehicles at no additional cost. Full Self-Driving (Supervised) — which adds city-street navigation, traffic light/stop sign response, and auto lane-change — requires either the original $8,000 one-time purchase or a $99/month subscription.
How does HW3 FSD compare to GM Super Cruise or Ford BlueCruise?
All three are SAE Level 2 systems requiring driver supervision. Tesla FSD on HW3 operates on a broader set of roads (any mapped street) but with less consistent performance and more frequent interventions. GM Super Cruise and Ford BlueCruise operate only on pre-mapped divided highways (~400,000+ and ~130,000+ miles respectively) but offer more reliable hands-free operation with robust driver-monitoring and, in GM's case, lidar-based map data. For highway-only use, Super Cruise and BlueCruise generally deliver a smoother, more predictable experience.
Where can I ride in a real self-driving robotaxi today?
Waymo One operates commercially in Phoenix (metro-wide), San Francisco (most of the city), Los Angeles (expanding zones), and Austin (limited launch area). Zoox has begun limited pilot operations in Las Vegas and Foster City, CA, but is not yet a commercial ride-hailing service. Cruise suspended operations after its October 2023 incident and has not resumed driverless commercial service as of 2026. Tesla's announced robotaxi service is not operational in any city.
