Estimating the Cost of Tesla HW4.0: Second-Generation FSD Options Could Approach $20,000

Greentheonly, a Tesla leak expert, has revealed detailed information about Tesla’s HW4.0, which allows us to estimate its cost.HW4.0 will first be used in the new version of Tesla’s Model X/S and is expected to be widely promoted in the future. Currently, the optional price for FSD in the U.S. is $15,000. In addition to the one-time hardware cost, there is also a monthly service fee of $199.

Image Source: FCC

HW4.0 includes one high-resolution millimeter-wave radar (estimated BOM cost $120), 11 camera modules (estimated BOM cost $240), and one computing box (for the intelligent driving part, estimated BOM cost $1075-1605). The total BOM cost is approximately $1500-2100, and this BOM cost is not fixed, as the cost of the critical FSD chip will fluctuate. The cost of the FSD chip will decrease rapidly with increased shipment volume, as the one-time wafer cost and R&D cost are allocated to each chip, and the higher the shipment volume, the lower the cost allocated to each chip. The software part’s cost may be higher, including data systems, data collection, organization, labeling, and training, which consume a lot of costs. However, it can be basically confirmed that the optional price for the second-generation FSD could approach $20,000.

Tesla reported the millimeter-wave radar to the FCC as early as June 2022. According to convention, internal disassembly should be announced by the end of 2022, but Tesla applied to the FCC for an extension until March 2023. However, greentheonly has already obtained internal disassembly images. Tesla’s previous generation millimeter-wave radar was the lowest-end product ARS-4B from Continental Automotive, with a procurement price of about 500-600 RMB, and a BOM cost not exceeding $50. The BOM cost of the millimeter-wave radar independently developed by Tesla on HW4.0 is about $120, and the overall cost is estimated to be $150-200.

Image Source: FCC

The PCB of the Tesla millimeter-wave radar has a shielding cover added to the low-frequency part. Excellent large manufacturers do not need to add shielding covers.

Image Source: FCC

Tesla used two 3-transmit and 4-receive radar transceivers in cascade, most likely Texas Instrument’s AWR2243. Previously, Tesla had already used Texas Instrument’s AWR6843 in the 60GHz radar inside the vehicle. Considering cost, Tesla did not use four transceivers in cascade like Continental’s ARS540, but only two in cascade, with each AWR2243 costing about $20. Tesla has a total of 48 channels, while Continental’s ARS540 has 192 channels.

Image Source: FCC

The PCB for the processing part of the Tesla millimeter-wave radar uses an FPGA from Xilinx, a subsidiary of AMD, model XA7Z020-1CLG400Q, which contains two ARM A9 cores, 130 thousand gates, 53,200 LUTs, and a DSP computing power of 276 GMACs, costing about $40-60. Domestic prices are relatively high, ranging from 900-1800 RMB. There is also a Micron DDR3 chip with a capacity of 125MB, model MT41K64M16TW-107 AUT, costing about $1. The upper part of the PCB may be an Ethernet PHY or MIPI CSI-2 PHY chip, and the right part is for power supply. The BOM cost is about $115, plus $5 for the PCB. This millimeter-wave radar is likely to be OEM by Quanta Computer in Taiwan.

Tesla HW4.0 Front PCB

Image Source: FCC

Greentheonly said on Twitter that the second-generation FSD’s CPU has increased from 12 to 20, with a frequency range of 1.37GHz to 2.35GHz. The first generation FSD had 12 ARM Cortex-A72 cores, in 3 clusters, with 4 cores each. The second generation is expected to still use A72, with 4 clusters, each with 5 cores. A72 is an architecture introduced by ARM in 2015. The main reason Tesla continues to use such an old architecture is likely due to cost. New ARM architectures can be very expensive, and only Rockchip and Huawei have purchased the A76 architecture domestically.

Second-Generation FSD Chip

Image Source: FCC

The cost of the FSD chip is difficult to estimate. According to the chip’s surface number, the second-generation FSD is likely still being manufactured by Samsung, whose quotes are only 1/3 or even 1/4 of TSMC’s, which is also the main reason TSMC’s profit margin is 5 times that of Samsung. Choosing Samsung makes it unlikely to choose an immature 5nm process, and it should still use a mature 7nm process. By mid-2023, Samsung’s Austin factory, located near Tesla, will be able to mass-produce 7nm chips.

The one-time wafer cost for 7nm chips is about $30 million, which is about 200 million RMB. If the second-generation FSD shipment reaches 100,000 units, the cost allocated to each chip is about $300; if the shipment reaches 300,000 units, the cost allocated to each chip is $100; if it reaches 1 million units, the cost allocated to each chip is about $30. Its Die size looks similar to NVIDIA Orin, and the estimated wafer manufacturing cost and packaging cost is $150. Tesla sold 1.31 million vehicles in 2022, with an FSD option rate of 20%, which means 300,000 sets. Assuming R&D costs of $9 million, the cost allocated to each chip is about $30, leading to a total cost of about $280; if the shipment volume is 150,000 units, the cost would be about $410; if it is 100,000 units, the cost would be $540.

Image Source: FCC

In addition to the two expensive FSD chips, the second most expensive component is the memory. Tesla has lavishly used 16 of the most expensive GDDR6 memory chips, FBGA code D9ZPR, actual model MT61M512M32KPA-14 AAT:C, with a capacity of 2GB, totaling 32GB.

GDDR (Graphics Double Data Rate) is a type of video memory specifically designed for high-end graphics cards, with dedicated operating frequencies, clock frequencies, and voltages, making it different from standard DDR memory on the market. Generally, it has a higher clock frequency and lower heat generation than ordinary DDR memory, making it more suitable for pairing with high-end display chips. GDDR is the high-end memory familiar to computer enthusiasts, and GDDR6 began to appear with NVIDIA’s 20 series graphics cards released in 2018. However, the most powerful consumer memory currently is GDDR6X, launched in 2020 by NVIDIA in collaboration with Micron.

GDDR6 has high power consumption and is generally only used in the computer field. Micron first introduced GDDR6, and by early 2021, the price of MT61M512M32KPA-14 AAT:C was about $14-16 per chip; currently, it is estimated at about $10 per chip, totaling $160 for 16 chips. It should be noted that the HW4.0 cockpit system also uses expensive GDDR6, but it is Samsung’s 4 pieces of 2GB DDR6, which is estimated to cost around $40. In terms of flash storage, HW3.0 used Toshiba’s THGAF8G8T23BAIL, which is a 32GB UFS but is an older UFS2.1 standard. HW4.0 switched to Samsung’s KLUDG8J1ZD, with a capacity increased to 128GB, but still using the UFS2.1 standard, costing about $7.5.

HW4.0 uses about 450 chips, most of which are for power management. In addition to the more expensive chips mentioned above, there is also an Ethernet switch chip in the center at the bottom. HW3.0 used Marvell’s 88Q6321, and HW4.0 will obviously not use this outdated non-strict automotive Ethernet chip. It is speculated that HW4.0 may have switched to the more advanced Broadcom BCM8956X or BCM8947X, or possibly Realtek from Taiwan, as the Ethernet switch on the vehicle board is from Realtek. Assuming it is the same Ethernet switch as the vehicle board, the price is about $30.

HW3.0 reportedly has 4746 components, while HW4.0 is estimated to have nearly 5000 components, most of which are passive components. Passive components are relatively low-cost, generally only a few cents in RMB, with the most expensive being tantalum capacitors, which are often advertised on high-end graphics cards or motherboards, and there are some in HW4.0, with tantalum capacitor prices ranging from 2-6 RMB. There are also 24 power supply paths using relatively advanced DrMOS, priced at about $1.3-1.7, which can only be seen on motherboards exceeding 2000. DrMOS is rapidly advancing domestically, with automotive-grade already in mass production, while common overseas brands are MPS and On Semiconductor.

Image Source: FCC

On the bottom left, there are three deserialization chips. According to the PCB markings, all three are 4-channel deserialization chips, meaning a maximum of 12 cameras, and Tesla has reserved positions for a few cameras, estimating that HW4.0 uses 8 cameras.

Image Source: FCC

On the left is the smart driving interface of HW3.0, and on the right is the interface of HW4.0. The upper layer is the smart driving interface of HW4.0, and the lower is the cockpit interface, which clearly has two display output interfaces. In terms of smart driving, the top red is reserved and currently unused. The blue is for driver behavior monitoring, and the four interfaces are for the 1.3 million pixel cameras on both sides of the vehicle. The white connects two rear cameras, and the black is for the front camera, which may both be 5 million pixels. It is speculated that Tesla has reduced one front-facing camera and added one rear camera. Tesla used three 4-channel deserialization chips, estimated to be Maxim’s MAX96712, which is very popular in China and can support 4 4 million pixel cameras, with a price of $35-70 each. HW3.0 used two Texas Instruments DS90UB960 and one DS90UB954, which are clearly smaller than Maxim’s. There should also be 11 additional serialization chips on the camera side, with an estimated total cost of $35.

There are 12 camera interfaces, of which 1 is reserved; 1 is for in-car driver behavior monitoring; 4 are for 360 surround view, i.e., “F-SVC”, “L-SVC”, “R-SVC”, “Back Up”, estimated to be relatively low 1.3 million pixels; 4 are for bird’s eye view BEV, i.e., “L-FF-Rear”, “R-FF-Rear”, “L-FF-Side”, “R-FF-Side”, FF is estimated to be far-field, about 2 million pixels; 2 are front-facing, one main and one wide field.

Data Source:Organized from public information

The above is HW4.0, and below is HW3.0, showing that the water cooling part has been redesigned

Image Source: FCC

HW3.0 is OEM by Quanta Computer’s Shanghai Songjiang factory, and HW4.0 is likely still OEM by Quanta Computer, but seems to have moved to OEM in Taiwan. Quanta is the world’s largest OEM manufacturer for laptops and servers, with revenue of about 288 billion RMB in 2022, and Tesla’s contribution to Quanta’s total revenue is less than 1%.

Tesla’s strongest point is still the application of water cooling. Only Tesla can use water cooling in the mass production vehicle field, which is also why Tesla dares to use desktop GPUs. Other manufacturers must master water cooling skills to compete with Tesla.

Disclaimer: This article only represents the author’s personal views.

More Zosi Reports

Contact for report ordering and cooperation consultation:

Mr. Fu: 15810027571 (same WeChat)

Zuo Yan Jun: 18600021096 (same WeChat)

Zosi’s 2023 Research Report Writing Plan

Panorama of the Intelligent Connected Vehicle Industry Chain (February 2023 Edition)

Domestic Brand OEM Autonomous Driving	Automotive Vision (Domestic)	High-Precision Maps
Joint Venture Brand OEM Autonomous Driving	Automotive Vision (Overseas)	High-Precision Positioning
ADAS and Autonomous Driving Tier1 – Domestic	Automotive Vision Algorithms	Automotive Gateway
ADAS and Autonomous Driving Tier1 – Overseas	Surround View Market Research (Local Edition)	Data Closed Loop Research
Key Components of ADAS Domain Controllers	Surround View Market Research (Joint Venture Edition)	Automotive Information Security Hardware
Autonomous Driving and Cockpit Domain Controllers	Infrared Night Vision	Automotive Information Security Software
Multi-Domain Computing and Regional Controllers	Autonomous Driving Simulation (Overseas)	OEM Information Security
Passenger Vehicle Chassis Domain Control	Autonomous Driving Simulation (Domestic)	Wireless Communication Modules
Domain Controller Ranking Analysis	LiDAR – Domestic Edition	Automotive 5G Integration
E/E Architecture	LiDAR – Overseas Edition	800V High-Voltage Platform
L4 Autonomous Driving	Core Components of LiDAR	Fuel Cells
L2/L2+ Autonomous Driving	Millimeter Wave Radar	Integrated Battery
Passenger Vehicle Camera Quarterly Report	Automotive Ultrasonic Radar	Integrated Die Casting
ADAS Data Annual Report	Radar Disassembly	Automotive Operating Systems
Joint Venture Brand Vehicle Networking	Ranking of Laser and Millimeter Wave Radar	Steer-by-Wire Research
Vehicle Information Service System and Entertainment Ecosystem	Dedicated Vehicle Autonomous Driving	Skateboard Chassis
Autonomous Driving Heavy Trucks	Mining Autonomous Driving	Electric Control Suspension
Commercial Vehicle ADAS	Unmanned Shuttle	Steering Systems
Commercial Vehicle Smart Cockpit	Unmanned Delivery Vehicle	Wire Control Brake Research
Commercial Vehicle Vehicle Networking	Unmanned Retail Vehicle Research	Charging and Replacement Infrastructure
Commercial Vehicle Smart Chassis	Agricultural Machinery Autonomous Driving	Automotive Motor Controller
Automotive Smart Cockpit	Port Autonomous Driving	Hybrid Power Report
Smart Cockpit Tier1	Modular Report	Automotive PCB Research
Cockpit Multi-Screen and Linked Screen	V2X and Vehicle Road Coordination	IGBT and SiC Research
Smart Cockpit Design	Roadside Intelligent Perception	EV Thermal Management System
Instrument and Central Control Display	Roadside Edge Computing	Automotive Power Electronics
Smart Rearview Mirror	Automotive eCall System	Electric Drive and Power Domain
Dash Cameras	Automotive EDR Research	Automotive Wiring Harness
Automotive Digital Key	Smart Automotive Personalization	Automotive Sound System
Automotive UWB Research	Automotive Multi-Modal Interaction	Automotive Seats
HUD Industry Research	In-Car Voice	Automotive Lighting
Human-Machine Interaction	Car Antennas	Automotive Magnesium Alloy Die Casting
In-Car DMS	TSP Manufacturers and Products	Denso’s New Four Modernizations
OTA Research	Autonomous Driving Regulations	New Forces in Vehicle Manufacturing – NIO
Automotive Cloud Service Research	Autonomous Driving Standards and Certification	Disassembly of NIO ET5/ET7 Intelligent Functions
Automotive Functional Safety	Intelligent Connected Testing Base	New Forces in Vehicle Manufacturing – XPeng
AUTOSAR Research	PBV and Automotive Robots	XPeng G9 Function Disassembly
Software-Defined Vehicles	Flying Cars	New Forces in Vehicle Manufacturing – Li Auto
Software Suppliers	Integrated Parking Research	Li Auto L8/L9 Function Disassembly
Passenger Vehicle T-Box	Smart Parking Research	Automotive VCU Research
Commercial Vehicle T-Box	Automotive Car-Sharing	Cockpit SOC
T-Box Ranking Analysis	Shared Mobility and Autonomous Driving	Automotive MCU Research
Model Supplier Research	Automaker Digital Transformation	Automotive MCU Research
DJI Front Dual Camera and Tudatong LiDAR Disassembly	Autonomous Driving Fusion Algorithms	Sensor Chips
Disassembly of NIO, Toyota, and Great Wall Vehicle Machines and Cockpit Domain Controllers	AI Large Models and Autonomous Driving Computing Centers	In-Vehicle Storage Chips
Smart Surfaces	Automotive CIS Research

「Zosi Research Monthly Report」

ADAS/Intelligent Vehicle Monthly Report | Automotive Cockpit Electronics Monthly Report | Automotive Vision and Radar Monthly Report | Battery, Motor, and Control Monthly Report | In-Vehicle Information System Monthly Report | Passenger Vehicle ACC Data Monthly Report | Front Vision Data Monthly Report | HUD Monthly Report | AEB Monthly Report | APA Data Monthly Report | LKS Data Monthly Report | Front Radar Data Monthly Report