MaiTech MEGA188 Edge AI Chip: Fully Domestic 3D Distributed TPU Architecture Breaks the Memory Wall

[Editor's Note] As an annual benchmark event in the semiconductor industry, the 8th IC Fengyun List is positioned at a critical window for industrial technology iteration and domestic substitution. This year's list has been fully upgraded, featuring 70 heavyweight awards covering eight core fields including AI, embodied intelligence, and smart vehicles, with multiple special awards newly added and optimized. The judging committee is composed of over 100 semiconductor investment alliance units and 500+ industry CEOs, and the winners will be announced at the 2027 Semiconductor Investment Annual Conference ceremony, striving to identify industry benchmarks and set up models of industry innovation.

[Candidate Enterprise] Shenzhen Maitexin Technology Co., Ltd. (abbreviated as "Maitexin")

[Candidate Award] Annual AI Technology Breakthrough Award

[Candidate Product] MEGA188 Edge-Side AI Inference Chip

In the era when large models and embodied intelligence are accelerating toward edge terminals, demand for real-time local large-model inference from robots, smart education, and industrial edge applications is growing explosively. However, traditional NPU architectures are severely constrained in edge-side inference by high power consumption and the "memory wall" bottleneck of data movement, and high-end accelerator chips and advanced-process supply chains face uncontrollable risks. Breaking through the limitations of separated storage and computing and building a fully domestic, autonomous and controllable, high-energy-efficiency edge-side large-model chip has become a core direction of attack for China's semiconductor industry in seizing the high ground of embodied intelligence and edge intelligent computing.

Facing this industry demand, Shenzhen Maitexin Technology Co., Ltd. (hereinafter referred to as "Maitexin"), which focuses on the R&D of edge intelligent computing chips, has launched a heavyweight edge-side AI coprocessor chip — MEGA188. As a national high-tech enterprise, a Guangdong Province/Shenzhen overseas high-level talent team enterprise, and a Shenzhen seed unicorn enterprise, Maitexin was founded under the leadership of Professor Yu Hao, founding dean of the School of Microelectronics at Southern University of Science and Technology, and its core team brings together senior experts from top institutions such as Huawei HiSilicon, ZTE Microelectronics, and UNISOC, with more than 70% holding master's or doctoral degrees.

MEGA188 abandons the traditional NPU design paradigm and pioneers a 3D distributed TPU architecture:

• Fully domestic mature process: Manufactured using domestic mature processes, with large-capacity 3D DRAM vertically stacked and integrated on-chip, physically shortening the data access path and alleviating the storage bottleneck of edge-side large-model inference;

• High energy efficiency and low power consumption: Achieves stable local large-model inference within a limited power budget, and can smoothly run mainstream large language and multimodal models under fanless passive cooling conditions, balancing energy efficiency with quiet reliability;

• Autonomous controllability and system adaptation: The chip has a built-in RISC-V open-source control core and integrates standard high-speed interfaces, enabling seamless collaboration with mainstream domestic main control platforms without relying on ARM or x86 ecosystem licensing, completely eliminating dependence on overseas IP and advanced processes.

Full-stack autonomous software toolchain and patent moat

To ensure absolute autonomous controllability of the underlying technology, Maitexin has independently developed for MEGA188 a neural network compiler for the RISC-V platform, a training-free hybrid quantization algorithm for large models, attention compression model processing, and distributed edge-side acceleration technology.

At the intellectual property level, the company has cumulatively applied for 22 patents around MEGA188 (including 21 invention patents), and has been granted 13 invention patents and 1 utility model patent. The patent layout comprehensively covers the entire chain of "3D distributed TPU architecture — neural network compilation — algorithm compression — embodied intelligence scenario applications," building a rigorous technical protection network.

Deploying distributed Token factories to turn communications infrastructure into Token productivity

Beyond edge-side large-model chips, Maitexin is extending its capabilities to the system level, expanding its "distributed Token factory" business — moving AI computing power traditionally concentrated in the cloud down to edge communications infrastructure such as base stations. Relying on micro base stations (RAN) and various edge communication nodes, edge large models are deployed through edge-side computing power cards equipped with Maitexin's MEGA188 chip, turning every node into a distributed token production node and enabling local inference, local token generation, and local fusion. With a low investment threshold and short deployment cycle, it can quickly cover distributed nodes across a city, transforming massive communications infrastructure into efficient Token productivity.

As the brain chip, MEGA188, together with MEGA088 (perception chip) and Maitexin's self-developed personal intelligent agent, will jointly support two major application scenarios:

Mobile terminal intelligent agents: Running personal intelligent agents on new AI hardware such as mobile phones, PCs, and tablets, consuming Tokens nearby within base station coverage areas and executing tasks, with interaction data flowing back locally to continuously iterate the token loop;

Embodied terminal intelligent agents: For embodied terminals composed of a "brain, cerebellum, and body," achieving end-to-end autonomous control from perception to decision execution — robot dogs going deep into special environments to complete power inspections, robots autonomously taking elevators to achieve contactless food delivery, and embodied intelligent agents integrating into daily companionship, all consuming Tokens on demand and operating networked under the coverage of distributed Token factories.

Breakthroughs in embodied intelligence and multi-scenario commercialization

With industrial-grade wide-temperature design, high energy efficiency, and privacy security that keeps data local, MEGA188 has achieved large-scale commercialization in multiple high-barrier scenarios:

Embodied intelligence and robotics: In robotic arm visual control scenarios, MEGA188 collaborates with domestic main controls to run vision-language models and diffusion models locally, compressing end-to-end latency from perception to control commands to an extremely low level; in the service robot field, it has been successfully introduced to leading service and humanoid robot manufacturers, solving the deployment challenges of elevator control and building delivery, and ensuring autonomous operation in weak-network or disconnected environments.

Smart education all-in-one devices: In education tablet and education notebook solutions, MEGA188 independently carries large language model inference locally, supporting intelligent Q&A, essay grading, and oral practice, ensuring that students' sensitive data does not leave the device.

From breaking through high power consumption and storage bottlenecks to achieving large-scale deployment of a fully domestic 3D distributed TPU architecture, Maitexin has filled the gap in domestic low-power edge-side large-model chips with hardcore technological breakthroughs. Looking ahead, Maitexin will continue to deepen the technological evolution of compute-storage integration and edge-side large-model chips, building a solid computing power foundation for the high-quality autonomous evolution of China's embodied intelligence and edge intelligent computing industries.

[Award Application Entry]

The 2027 Semiconductor Investment Annual Conference and IC Fengyun List Award Ceremony will be held in December 2026. Award applications have been launched, and reporting on solicited and candidate enterprises/institutions is currently underway. Welcome to sign up and participate, and join us at this industry feast!

[Annual AI Technology Breakthrough Award]

Aimed at honoring enterprises or research institutions that in 2026 achieved major original innovations in AI fundamental theory, core algorithms, or key application fields, with technology reaching internationally leading levels, and that are of great significance to promoting the autonomous innovation and industrial development of China's AI technology.

[Application Conditions]

Achieving major technological breakthroughs in a certain subfield of artificial intelligence (such as large models, computer vision, natural language processing, etc.);