As large models transition from the virtual world to the physical space, embodied AI has become a new focal point in the global technological competition.
According to forecast data from the Gaogong Robot Industry Institute (GGII), global sales of humanoid robots will experience explosive growth in 2026, approaching 94,700 units, achieving a nearly fivefold leap compared to 18,500 units in 2025.
In this globally anticipated wave, a crucial hardware fact often overlooked by the public is that up to 90% of the global humanoid robot supply chain is rooted in China. Moreover, in the bill of materials (BOM) cost of the whole machine, the drive and execution systems responsible for motion (joints and dexterous hands) account for more than half.
However, to enable robots to flexibly and smoothly perform delicate tasks such as serving tea and water, or factory assembly, the underlying microcontrollers (MCUs) are facing unprecedented technical challenges.
Li Yi, Product Marketing Director of the MCU Business Unit at GigaDevice, told eefocus journalists that traditional industrial automation boards are bulky, often requiring a discrete layout of 5 to 7 chips, including MCUs, FPGAs, communication slave controller chips (ESC), and multiple independent physical layer transceivers (PHY). However, when attempting to cram this system into the highly space-constrained toroidal joint cavity of a humanoid robot, strict spatial limitations eliminate the possibility of convective heat dissipation for the solid-state hardware boards. Meanwhile, improvements in robot joint motion control performance lead to a surge in dynamic power consumption, while miniaturized packaging drastically reduces the heat dissipation area. This extreme imbalance between heat generation and dissipation is highly prone to triggering thermal runaway, ultimately causing the collapse of the entire motion control system.
Facing this industry pain point that hinders scaled industrial implementation, motion control chips are accelerating their shift from "general-purpose products" to "deeply customized solutions for specific scenarios".
As a leader in domestic general-purpose microcontrollers, GigaDevice, with its keen insight to secure a position in the blue ocean of intelligent motion control, has taken the lead in launching the GD32H77R series of high-performance MCUs dedicated to robots and the GD32F50MxxG series of highly integrated motor control MCUs. These aim to reshape the design boundaries of robot servo hardware with a breakthrough "single-chip solution" approach.
Against this backdrop, eefocus journalists interviewed Li Baokui, Senior Vice President of GigaDevice and General Manager of the MCU Business Unit, and Chen Siwei, Senior Director of Product Marketing for the MCU Business Unit, engaging in an in-depth dialogue on the most concerning technical routes and commercial competition in the industry.
Scale Effects Emerge: Domestic Chips Have Occupied the Main Track
Currently, the humanoid robot industry is in a stage of rapid scaling. The overall market size has not yet reached mass volume, and costs must be driven down through scale. So why would GigaDevice abandon reusing general-purpose chips from industrial categories and instead develop two new robot-dedicated MCUs at this time?
In response, Li Baokui frankly stated, "When the industry was just starting, the market size was too small. Investing heavily in dedicated chips would indeed make the ROI uncalculable, so everyone used industrial general-purpose chips as a stopgap. But now, the technological direction and shipment volume of embodied AI have reached a significant turning point. Robot joint cavities require small size and high-temperature resistance. These special requirements are far different from traditional industrial products, necessitating a dedicated and differentiated route."
"While the industry is still discussing when humanoid robots can truly achieve scaled mass production and move towards applications, as a chip manufacturer, GigaDevice has actually been steadily consolidating its foundation. In the first half of 2026, GigaDevice's MCU shipments in the embodied AI field exceeded 3 million units. Because it captured more than half of the market share in the early stage of the track, this massive base has brought sufficient scale effects."
To Achieve Hardware Decoupling, GD32H77R Opts for Ultimate Single-Core Performance
As system complexity increases, dual-core (e.g., M7+M4) or multi-core MCUs are flooding the market. Why, then, does GigaDevice's newly promoted flagship product, the GD32H77R, insist on a "single-core" architecture?
Chen Siwei, Senior Director of Product Marketing for the MCU Business Unit at GigaDevice, provided a relatively in-depth technical clarification on this. He explained that traditional industrial automation (such as high-precision CNC machine tools) indeed requires multi-core coordination to handle particularly complex multi-axis motion and high-precision machining. However, the joints of humanoid robots are basically single-axis controlled and do not require such high control precision at all.
In fact, in the joints of humanoid robots, a dual-core architecture is rather a burden. First, dual cores generate more heat within the chip itself, which is highly prone to triggering thermal runaway in compact cavities without convective heat dissipation. Second, multi-core designs bring additional power consumption losses and cost increases.
More crucially, the hardware design of the GD32H77R cleverly leverages the lightweight characteristics of the EtherCAT® bus. Conventional industrial bus protocols indeed require a second core to run the communication protocol stack. However, the GD32H77R integrates a Beckhoff-licensed EtherCAT® slave controller and two 100M Ethernet PHYs on-chip. The underlying hardware network core automatically completes all the data unpacking from the network cable within the chip. Once unpacked, the data is directly placed into the process data RAM by the hardware, and the main core (CPU) only needs to read it directly, without consuming any overhead for the communication protocol stack, achieving true "zero CPU overhead."
Chen Siwei stated frankly that with a single-core 600MHz M7 core, coupled with tightly coupled memory (TCM) running at the same frequency as the CPU, interrupt response and real-time motor control have been optimized to the extreme. A dual-core architecture is not a rigid demand for humanoid robot joints.
Additionally, although the GD32H77R, targeted at joint modules, boasts strong performance, core voltage optimization reduces the core voltage to only about 65% of that used by competitors. Combined with upgrades to the process platform, the chip's dynamic power consumption and self-heating are significantly reduced. Coupled with its compact 9mm×9mm BGA169 package, the volume is directly reduced to 1/3 of the previous generation, effectively solving the triangular dilemma of "performance, temperature, and size."
Ultimate 3-in-1 Integrated Packaging: GD32F50MxxG Empowers Micro Joints
If the GD32H77R is the "thigh joint main controller" responsible for high dynamic response, then the GD32F50MxxG series is the "preferred choice for upper limb and dexterous hand drive," focusing on high cost-effectiveness and ultimate integrated packaging.
It is equipped with an Arm® Cortex®-M33 core with a maximum clock frequency of 252MHz and adopts SiP (System in Package) technology. Within a very small PCB layout area, it integrates a high-performance microcontroller, a three-phase gate driver (pre-driver) with high driving capability, and a 4-channel rail-to-rail Op-Amp (Operational Amplifier) into a single package, creating a "3-in-1" single-chip solution.
Regarding the market competitiveness, positioning boundaries, and cost issues of the integrated packaging solution, both Li Yi and Chen Siwei provided technical clarifications. They believe that most existing integrated or all-in-one chips on the market are concentrated in mid-to-low-end applications (such as power tools), with built-in SRAM often only 8KB or 16KB, which is completely insufficient to run the complex closed-loop control and high-frequency task parallel scheduling algorithms of humanoid robots. In contrast, the GD32F50MxxG is a hardcore product that truly combines a high-performance computing core, a high-driving-capability pre-driver, and high-end analog Op-Amps. It comes standard with up to 1MB Flash and 128KB SRAM (of which 32KB features ECC verification), providing sufficient computing power to easily achieve a dual-axis 40kHz switching frequency.
There are also targeted designs in the details. High-frequency electromagnetic noise is generated during motor operation. The Op-Amps integrated in the GD32F50MxxG feature built-in high-performance RF/EMI filters, which can effectively suppress high-frequency noise, creating a clean and stable electrical environment for 12-bit ADC sampling and ensuring sufficiently accurate current sampling signals. Meanwhile, the built-in three-phase pre-driver has a high-side voltage tolerance of up to 120V, with peak driving capabilities of 2A source current and 2.5A sink current. It also features built-in bootstrap diodes, allowing it to directly drive peripheral power devices. This deeply integrated solution can effectively handle voltage spike impacts caused by motor back-EMF, prevent power stage damage, and enhance system robustness under harsh operating conditions.
Regarding pricing strategy and cost control, Chen Siwei added: Currently, the global semiconductor supply chain, from wafer fabs to packaging plants, is seeing price increases, and raw material costs continue to rise. However, from the perspective of the comprehensive cost of the whole machine, the advantages of the integrated solution are still unmatched by discrete solutions.
In traditional discrete solutions, adding an independent ESC chip to a high-performance MCU costs an extra two to three US dollars. Coupled with 2 to 4 independent Op-Amps and gate driver circuits, the total PCB area occupied is two to three times that of the GD32F50MxxG, and the routing is complex and troublesome. Although the GD32F50MxxG uses more expensive high-thermal-conductivity substrates, special molding compounds, and high-end substrates within its package to solve internal heating issues—making the cost of a single chip slightly higher—it directly saves whole machine customers the material costs of multiple peripheral chips and cuts the driver board PCB area in half. More importantly, once the chip size is reduced to a certain extent, it solves the critical bottleneck of "whether it can fit into the toroidal joint cavity," rather than just being a few dollars cheaper.
Therefore, this highly integrated single-chip solution is accelerating the elimination of traditional discrete component solutions with structural advantages.
In terms of package selection, the GD32F50MxxG is also highly targeted, offering two miniaturized models: GD32F50MMGO7G (8mm×8mm QFN80) and GD32F50MVGK7G (7mm×7mm BGA100). The QFN80 package is more SMT soldering-friendly, suitable for space-constrained medium-power applications. The BGA100 package, with more ball grid array solder joints, offers better thermal performance through high-density interconnects, making it the preferred solution for harsh, high-power operating conditions.
"Compliant Global Expansion + Hardcore Ecosystem" Are the Two Main Pillars for the Commercial Growth of Domestic Chips
The chip hardware architecture is the "muscle" and "brain," but the software ecosystem, information security, and compliance for global expansion are the keys to determining whether the chip can truly achieve large-scale implementation.
Li Baokui mentioned that embodied AI is a brand-new track where everyone is starting from scratch at the same starting line. In the past, traditional automotive or industrial fields might have been absolutely dominated by overseas brands. But on this new track, Chinese chip manufacturers, whether in the richness of technical routes, the closeness to the pain points of frontline customers, or iteration efficiency, have already run ahead of overseas manufacturers.
In terms of information security, humanoid robots will enter factories in large numbers in the future, and even homes. The requirements for information security and functional safety are even higher than those for automobiles. Both the GD32H77R series and the GD32F50MxxG series are equipped with complete security modules, including a TRNG (True Random Number Generator), a SHA256 hash unit, and AES128/192/256 hardware encryption engines.
More crucially, GigaDevice is the first domestic MCU manufacturer to truly thoroughly understand and strictly comply with the EU Cyber Resilience Act (CRA) and IEC62443 security standards. From chip architecture design and underlying firmware to long-term lifecycle maintenance, all comply with EU regulatory requirements. This lays the most fundamental security and regulatory defense line for downstream whole machine OEMs to sell robots overseas and open up the European market.
As of the first half of 2026, GigaDevice's MCU shipments in the embodied AI industry have exceeded 3 million units, capturing more than half of the market share, serving over 300 upstream and downstream robot customers globally.
Conclusion
From the simple migration of general-purpose chips to deep customization tailored specifically for robot joints and single-chip one-stop integration, this motion control technology upgrade driven by Chinese "chips" is building the most solid and highly energy-efficient hardware foundation for the lightweighting and comprehensive industrial implementation of embodied AI.