In the past, discussions about LiDAR focused more on lasers, detectors, and scanning mechanisms. Today, an increasing number of manufacturers are designing their own custom ASICs and SoCs (System on Chip), enabling chips to take on a growing share of signal and data processing tasks.
So, what exactly are these chips inside LiDAR processing?
01. Why Doesn't LiDAR Receive Ready-Made Point Clouds?
What LiDAR actually receives is reflected light. After the photodetectors convert these optical signals into electrical signals, the system must still undergo processes such as amplification, filtering, comparison, analog-to-digital conversion, or time measurement to obtain data usable for ranging. Taking the common ToF (Time-of-Flight) LiDAR as an example, its transmitter emits laser pulses, which are reflected by the target and return to the detector. The receiving end needs to determine whether a target exists from the extremely weak echo signal and ascertain the arrival time of the echo.
During this process, the analog front-end is responsible for amplifying and conditioning the signals output by the detector. The ADC (Analog-to-Digital Converter) converts waveforms into digital data, while the TDC (Time-to-Digital Converter) directly measures time-related information. Subsequent digital processing is then handled by an FPGA, ASIC, or SoC. In other words, the point cloud is the result of this series of steps. What the chips actually do is convert raw electrical signals step by step into measurement information such as distance and reflectivity, rather than directly calculating the point cloud.
The architectures of different products are not entirely identical. Some adopt multiple dedicated components, while others complete digital processing via FPGAs or other processors. An increasing number of products are integrating functions such as data acquisition, time measurement, and point cloud processing into dedicated SoCs. Notably, current LiDAR systems are migrating from analog architectures to digital ones.
Traditional analog solutions rely on multiple discrete components, whereas the digital architecture centers on the SPAD-SoC, integrating photon reception, signal processing, and large-scale digital computing units into a single chip to achieve a fully digital link where photon incidence directly generates digital pulses. In this way, the starting point of the processing link is digitized; the SPAD converts individual photons directly into digital signals, bypassing the need for a complete analog amplification and ADC sampling link. Qiu Chunchao, CEO of RoboSense, analogizes this shift to the underlying transition from CCD to CMOS in the imaging industry, believing that digitization enables LiDAR to possess Moore's Law-like evolutionary capabilities for the first time.
Looking at the entire link, the chips essentially take on the task of progressively converting the raw electrical signals obtained by the detectors into usable measurement information such as distance and reflectivity. As for the final output format and the specific stage where the point cloud is formed, it depends on the specific LiDAR architecture.
02. Why Can't All This Processing Be Handed Over to the Vehicle's Autonomous Driving Chips?
Since the vehicle itself already possesses one or even multiple high-computing-power autonomous driving SoCs, why place dedicated processing chips inside the LiDAR? One reason is that LiDAR deals with data very close to the sensor's underlying layer. The raw signals generated by the detectors contain a massive amount of information that needs to be processed locally in real time.
If as much raw data as possible is transmitted to the vehicle for unified processing by the autonomous driving computing platform, it would require higher data bandwidth and increase transmission and computing burdens. Moreover, the LiDAR itself knows best how to extract valid information from its own optoelectronic signals. Placing part of the processing inside the sensor allows filtering, conversion, and compression to be completed before the data enters the vehicle's central computing platform, reducing unnecessary data transportation.
However, this does not mean there is an absolute boundary between LiDAR chips and autonomous driving chips. The vehicle's autonomous driving SoC focuses more on multi-sensor information fusion and higher-level tasks such as environment understanding, target prediction, decision-making, and planning; whereas the computing inside the LiDAR revolves more around its own signal acquisition and sensor data processing.
It can be simply understood that the LiDAR internals focus more on what is measured, while the vehicle's autonomous driving computing focuses more on what this information means and what the vehicle should do next. Nevertheless, as the integration level of LiDAR SoCs increases, this boundary is becoming increasingly blurred. Some new LiDAR SoCs are no longer limited to simple signal acquisition but are beginning to take on more complex tasks such as point cloud processing.
03. Why Are More and More Manufacturers Starting to Develop Custom ASICs and SoCs Now?
In the past, LiDAR could complete various stages such as signal acquisition, analog conditioning, analog-to-digital conversion, time measurement, digital processing, scanning control, and data output using general or dedicated components like FPGAs, ADCs, TDCs, analog front-ends, laser drivers, scanning drivers, and interface chips. However, as LiDAR evolves toward longer ranges, higher resolutions, higher point densities, and faster scanning rates, the data processing volume continues to increase. If it continues to rely on a large number of discrete components, the system will require more chips, larger PCB (Printed Circuit Board) space, and more complex data transmission, while also bringing pressure on power consumption, cost, and system design.
Custom ASICs and SoCs, on the other hand, perform hardware-level optimizations tailored to LiDAR data flows and processing tasks, further integrating functions originally handled by multiple components. In 2026, domestic manufacturers are also moving in this direction. RoboSense and Hesai Technology have successively released new-generation chip platforms, shifting the focus of competition from scanning mechanisms to the chips themselves. RoboSense launched the Phoenix SPAD-SoC, natively integrating 2,160 channels on a single chip with a maximum detection range of 600 meters; Hesai Technology introduced the Picasso 6D full-color SPAD-SoC and already has the Fermi C500 main control chip based on the RISC-V architecture. This does not mean LiDAR lacked chips in the past; rather, chips are transitioning from traditional electronic support components to crucial elements influencing LiDAR system architecture.
04. Will LiDAR Become Increasingly Dependent on Its Own Computing Power?
Judging from current technological routes, this trend is already quite clear, but it does not mean LiDAR is going to become an independent autonomous driving computer. The Autonomous Driving Frontier believes that a more accurate description is that LiDAR is gradually transforming from a sensor centered on optical and optoelectronic components into an intelligent sensor with stronger local computing capabilities. Hesai Technology proposes that the industry is leaping from spatial perception to spatial intelligence, while RoboSense positions LiDAR as a physical AI (Artificial Intelligence) data entry point, shifting the competitive focus from point clouds to pixels. In the future, LiDAR competition may not simply be about adding a more powerful chip, but rather continuing to improve the integration among light sources, detectors, analog front-ends, data acquisition, time measurement, and point cloud processing, enabling more computations highly related to the sensor's own characteristics to be completed locally.
For the vehicle, this approach can reduce data transportation and system burden; for LiDAR, it presents an opportunity to further shrink its size, lower power consumption and costs, while enhancing real-time processing capabilities. Therefore, when discussing the chips inside LiDAR today, what truly deserves attention is not whether there are chips, but what the chips are taking on, and why an increasing number of originally dispersed functions need to be integrated into a single dedicated chip. In the future, when evaluating a LiDAR system, apart from the laser, detector, scanning mechanism, and detection range, the chip architecture itself will also become increasingly worthy of attention.