AI Smart Glasses Chip Solutions: Deep Research Report

Industry Overview: The Paradigm Shift from Optical Display to Computational Intelligence

With the explosive growth of Generative AI between 2024 and 2025, the smart glasses industry is undergoing a profound paradigm shift. For a long time, the development of Augmented Reality (AR) glasses was constrained by the "impossible triangle" of optical display technology (such as waveguide efficiency and MicroLED mass production), physical form factor (weight and heat dissipation), and battery life.

However, the success of a new generation of AI glasses, epitomized by the Ray-Ban Meta, has demonstrated a shift in market focus. Users are no longer purely chasing complex holographic displays; instead, they are seeking "AI-native" terminals capable of all-day perception, instant information interaction, and first-person content creation. This shift moves the core competitive elements from optical modules to computing platforms, specifically focusing on the Performance-per-Watt ratio, multimodal perception capabilities, and the efficiency of On-Device AI deployment.

This report provides an exhaustive technical analysis of current and emerging chip solutions, ranging from Qualcomm’s high-performance heterogeneous platforms to Bestechnic’s cost-effective MCU+NPU architectures, alongside differentiated layouts from MediaTek, Unisoc, Rockchip, and Realtek. We will also examine the synergistic evolution of Image Signal Processors (ISP) and Neural Processing Units (NPU) in the era of on-device Large Language Models (LLMs).

AI Glasses Computing Challenges and Thermal Design Power (TDP) Constraints

Unlike smartphones, the physical form of smart glasses dictates extremely rigorous engineering constraints. To ensure comfortable all-day wear, the target weight for the device is 50–80g, which limits battery capacity to between 150mAh and 500mAh. Consequently, the sustained Thermal Design Power (TDP) budget for the computing unit is usually in the hundreds of milliwatts (mW). Peak power rarely exceeds 1-2 watts to prevent the temples from overheating and causing discomfort to the wearer's skin.

In this restricted power envelope, chip architectures must shift from general-purpose CPUs to highly specialized heterogeneous computing. The CPU manages logic control, while computationally intensive tasks—such as Real-time Speech-to-Text (ASR), Visual SLAM, image enhancement, and LLM inference—are offloaded to dedicated cores like NPUs, DSPs, and ISPs. This "dedicated cores for dedicated tasks" philosophy is fundamental to modern AI glasses chip design.

Qualcomm: Defining High-End Standards with Heterogeneous Computing

Qualcomm has leveraged its expertise in mobile SoCs to build the industry's most comprehensive line of dedicated XR/AR chips, establishing reference standards for high-end AI glasses with its Snapdragon Spaces ecosystem and hardware synergy.

Snapdragon AR1 Gen 1: The Foundation of Lightweight AI Glasses

Designed for glasses without displays or with lightweight Head-Up Displays (HUD), this platform powers hits like the Ray-Ban Meta by balancing flagship performance with sleek design.

Core Architecture and Process: Manufactured on a 4nm to 6nm process, the SoC features a quad-core Qualcomm Kryo CPU (up to 1.9GHz) capable of running Android or customized Linux.

Spectra ISP (Computational Photography): Includes dual 14-bit ISPs supporting 12MP photos and 6MP video. Key hardware-accelerated features include:

Computational HDR: Crucial for high-contrast outdoor scenes where physical aperture adjustment is impossible.

Hardware-level EIS: Uses six-axis electronic stabilization based on gyroscope data to ensure stable first-person video.

Real-time Perception: Hardware support for face detection to optimize exposure and white balance.

Hexagon NPU: The "soul" of the AR1, featuring scalar, vector (HVX), and tensor accelerators. It handles microphone beamforming, Echo Cancellation (AEC), and wake-word detection, while increasingly managing visual tasks like object recognition.

Connectivity: Supports Wi-Fi 7 (up to 5.8Gbps) and Bluetooth 5.3/5.4 via the FastConnect 7800 series, enabling low-latency cloud interaction and high-bandwidth streaming.

Snapdragon AR1+ Gen 1: A Catalyst for On-Device LLM

Released in 2024, the AR1+ focuses on on-device Generative AI.

Optimization: Achieves a 26% reduction in package size and a 7% reduction in power consumption.

On-Device LLM Support: Native support for Small Language Models (SLM), such as Meta Llama 3.2 1B. Optimized memory bandwidth and NPU units allow for millisecond-level intent understanding and multi-turn dialogue without an internet connection.

Snapdragon AR2 Gen 1: The Distributed Architecture Revolution

To solve the heat and wiring limits of complex AR glasses, the AR2 utilizes a revolutionary distributed architecture.

Multi-Chip Topology: Splits the SoC into three physical modules to balance weight and heat:

AR Processor (4nm): Located in the temple, it handles SLAM, gesture recognition, and 6DoF tracking for up to 9 concurrent cameras.

AR Co-processor: Located in the bridge, it manages eye-tracking and iris authentication. This reduces wiring through the hinge by 45%, improving mechanical reliability.

Connectivity Platform: Offloads heavy rendering to a host (phone) via Wi-Fi 7 with <2ms latency, keeping glass-side power below 1W.

Table 1: Snapdragon AR1 Gen 1 vs. AR2 Gen 1 Comparison

Bestechnic (BES): A Cost-Effective Heterogeneous Breakthrough Built on an RTOS Ecosystem

In the Chinese market and the mid-range AI smart glasses segment, Bestechnic (BES) has carved out a technological path distinctly different from Qualcomm’s. Leveraging its dominant position in TWS earbud chips, BES has pursued a core strategy centered on the “intelligent upgrade of audio SoCs.” By combining low-power MCU architectures with RTOS-based systems, the company delivers AI smart glasses solutions that emphasize extreme cost efficiency and long battery life.

BES2800: The Convergence of Audio and AI

The BES2800 is Bestechnic’s new-generation flagship SoC designed for smart wearable devices and is widely adopted in AI smart glasses projects by companies such as Alibaba (Tmall Genie) and Xiaomi. Rather than being a simple Bluetooth chip, it is an edge AI platform integrating heterogeneous multi-core computing capabilities.

Heterogeneous Multi-Core Architecture

Manufactured using an advanced 6nm FinFET process, the BES2800 integrates a complex heterogeneous subsystem:

● CPU Subsystem: It includes dual Arm Cortex-M55 cores, Arm’s low-power embedded AI processors that support Helium vector processing technology. In addition, dual-core STAR-MC1 processors are integrated as the host subsystem, responsible for system scheduling and application logic.

● BECO NPU: The chip integrates a dual-core BECO (Bestechnic Neural Network Co-processor) NPU. While its absolute compute power cannot rival that of smartphone SoCs, it is deeply optimized for specific audio and lightweight vision models. Compared with the previous-generation BES2700, the BES2800 delivers a significant NPU performance uplift, enabling efficient execution of keyword spotting (KWS), call noise cancellation (ENC), and sensor fusion algorithms.

● Integrated Connectivity: Dual-mode Bluetooth 5.4 and low-power Wi-Fi 6 are integrated on a single chip. This high level of integration significantly reduces BOM cost and PCB footprint, making it a key enabler for AI smart glasses priced at the RMB 1,000 level.

The “BES2800 + External ISP” Dual-Chip Architecture

The BES2800 does not natively include a powerful image signal processor (ISP) and therefore cannot directly handle complex processing tasks for high-resolution cameras, such as HDR and image stabilization. To compensate for this limitation, the industry—including companies like ByteDance and Looktech—commonly adopts a “BES2800 + discrete ISP” dual-chip architecture.

Architectural Logic and Task Allocation

In this architecture, system tasks are clearly divided:

● BES2800 (Main Controller): Responsible for running the RTOS, Bluetooth/Wi-Fi connectivity, audio processing (microphone array noise reduction and speaker output), touch interaction, and power management. It maintains ultra-low standby power consumption (at the microwatt level), ensuring extremely long standby time when the camera is inactive.

● Discrete ISP (Co-Processor): Dedicated to camera image acquisition, preprocessing (3A, denoising), video encoding (H.264/H.265), and selected visual AI tasks. The ISP is activated only when the camera is in use and immediately enters sleep mode once recording ends.

Key ISP Partners: Sigmastar and SuperAcme

● Sigmastar (SigmaStar) SSC309QL: A custom ISP chip designed specifically for smart glasses. It uses chiplet packaging technology with integrated LPDDR4x memory, significantly reducing package size. The SSC309QL supports 12MP image capture and 4K video encoding and integrates 1.5 TOPS of AI compute for on-device image recognition. Its power efficiency is outstanding: recording at a 2 Mbps bitrate consumes only about 300 mW, keeping total system power around 600 mW—roughly 50% of Qualcomm’s AR1 solution.

● SuperAcme: Another domestic ISP supplier whose chips were originally designed for security IPC (network cameras) and later adapted for smart glasses. With exceptional cost efficiency, SuperAcme enables manufacturers to reduce AI smart glasses BOM costs to below RMB 1,000, allowing retail prices to drop to under RMB 1,500.

BES2700: A Mature Co-Processor Role

In earlier products (such as Huawei smart glasses) or certain high-end dual-chip configurations, the BES2700 is often used as a co-processor. For example, in some Xiaomi AI smart glasses configurations, a “Qualcomm AR1 (vision processing) + BES2700 (audio and ultra-low-power standby)” big–little system-level division of labor may be employed. This design exploits BES chips’ absolute advantage in audio power efficiency: the BES chip handles daily Bluetooth audio streams, while the Qualcomm chip is only activated for photography or large-model inference, balancing performance and battery life.

MediaTek and Airoha: Strategic Challengers Poised to Rise

Rather than aggressively launching XR-specific platforms early, MediaTek has adopted a more discreet dual-track strategy of “high-end customization + subsidiary coverage.” As its strategic collaboration with Meta deepens, MediaTek is emerging as the most formidable potential challenger in the high-end market.

Deep Customization with Meta

It is widely believed that Meta and MediaTek are collaborating on the development of next-generation AR SoCs, although specific details have not been publicly disclosed. This move is seen as a key strategy for Meta to reduce its reliance on Qualcomm and gain greater control over underlying hardware definitions.

● Technical Path: Leveraging MediaTek’s accumulated expertise in high-efficiency APUs (AI processing units) from the Dimensity series and advanced connectivity technologies such as Wi-Fi 7, the goal is to develop SoCs with extreme energy efficiency to address the “compute–power–form factor” trilemma of AR glasses.

● Future Products: Industry speculation suggests that Meta’s next-generation smart glasses with displays—codenamed “Hypernova” or “Celeste”—are likely to adopt MediaTek custom chips or a hybrid architecture combining a Qualcomm primary chip with a MediaTek co-processor. Leaked information indicates the inclusion of a micro-display (HUD), which places higher demands on display driving and low-power rendering capabilities.

Airoha: The Hidden Champion of Audio Glasses

As a MediaTek subsidiary, Airoha holds a strong position in the Bluetooth audio SoC market. Its AB1562/AB1565 series, although primarily designed for TWS earbuds, is widely used in entry-level audio smart glasses.

● AB1562 Series: Supports Bluetooth 5.3 and LE Audio and integrates a Cadence HiFi DSP. While lacking video processing capabilities, it delivers industry-leading performance in ANC, echo cancellation, and ultra-low-power audio playback. For camera-free, voice-centric AI glasses (such as ChatGPT-based voice interaction glasses), this represents a highly mature and extremely low-cost solution.

● Software Support: Supports MCSync technology for synchronized dual-ear transmission with reduced latency and provides a complete SDK enabling integration with third-party voice assistants such as Alexa and Google Assistant.

Unisoc, Rockchip, and Realtek: Optimal Choices for Differentiated Scenarios

Unisoc W517: Turning Android Watches into Glasses

Originally designed as a flagship smartwatch platform, the W517 has been widely adopted in Android-based smart glasses such as Inmo Air and Sharge Loomo due to its full standalone connectivity (4G/LTE) and solid compute performance.

● Architectural Advantages: Built on a 12nm process with a quad-core CPU (1× Cortex-A75 + 3× Cortex-A55). This “small-core–centric” design suits thermally constrained devices. Most importantly, it integrates a 4G baseband, enabling glasses to operate independently with a SIM card for calls and data, without constant reliance on a smartphone—an attractive feature for outdoor sports and standalone navigation scenarios.

● AI and Multimedia: Supports AI noise reduction, image stabilization, and face recognition. The integrated IMG8300 GPU enables 1080p video encoding/decoding. While less powerful than AR1, the flexibility of the Android ecosystem allows extensive application expansion through software.

Rockchip RK3588: A Compute Powerhouse for Edge Computing and Heavy AR

The RK3588 is a high-performance general-purpose SoC. Although its power consumption (TDP 5–10 W) is too high for lightweight glasses, it is commonly used in tethered AR glasses compute boxes or development and validation platforms.

● Extreme NPU Performance: Manufactured on an 8nm process, it features an octa-core CPU (4× A76 + 4× A55) and up to 6 TOPS of NPU compute. This enables smooth local execution of 1B or even 3B parameter LLMs (e.g., TinyLlama, MiniCPM), with typical token generation speeds of 3–8 tokens/s and up to 10–15 tokens/s under extreme optimization—well above human reading speed.

● Industrial-Grade Vision: Supports 8K video encoding/decoding and concurrent multi-camera input. Its NPU supports INT4, INT8, and FP16 precision, and when paired with the RKNN toolchain, delivers highly efficient model deployment—ideal for industrial AR applications such as remote assistance and complex object recognition.

Realtek RTL8763E: Low-Power Audio-Visual Integration

The Realtek RTL8763E series (e.g., RTL8763EWE) integrates Bluetooth audio, an MCU, and basic video processing capabilities, making it suitable for entry-level camera-equipped glasses.

● Key Features: Supports dual-mode Bluetooth 5.3, integrates a high-performance DSP and AI NN engine, and enables keyword spotting (KWS) and basic image recognition. It represents a middle-ground solution—more capable than pure audio chips and significantly cheaper than Qualcomm—well-suited for cost-sensitive cross-border e-commerce products.

On-Device AI Software Ecosystem and Model Deployment

Hardware defines the lower bound of AI smart glasses, while the software stack and model deployment capabilities define the upper bound of user experience. With the release of lightweight models such as Llama 3.2, on-device AI has become a key battleground among chip vendors.

Qualcomm AI Stack and Snapdragon Spaces

Qualcomm has built a comprehensive ecosystem from low-level drivers to application layers:

● Snapdragon Spaces: An open XR developer platform supporting Unity and Unreal Engine. It provides standardized APIs that allow developers to easily access perception algorithms such as plane detection and hand tracking without dealing with hardware specifics. Spaces is actively migrating toward Android XR standards to achieve cross-device compatibility.

● AI Model Deployment: Qualcomm AI Stack supports converting PyTorch and TensorFlow models into the QNN (Qualcomm Neural Network) format. The AR1+ Gen 1 supports INT4/INT8 quantized execution of Llama 3.2. Through heterogeneous computing strategies—CPU for preprocessing and NPU for dense tensor computation—low-latency, low-power inference is achieved. Qualcomm also collaborates with Google to further improve NPU efficiency through the LiteRT (formerly TensorFlow Lite) accelerator.

BES and AI Challenges Under RTOS

For RTOS-based platforms such as the BES2800, AI deployment faces greater challenges due to the lack of rich libraries comparable to Android.

● Lightweight Models: Bestechnic provides toolchains optimized for its BECO NPU, supporting TensorFlow Lite Micro model conversion and execution on DSP/NPU. However, this is typically limited to tasks such as keyword spotting and simple classification.

● Cloud Collaboration: Due to limited on-device compute, RTOS-based glasses often adopt a “device-side detection + cloud inference” model. The on-device NPU handles voice activity detection (VAD) or trigger events, then transmits data to a smartphone via Bluetooth. The smartphone app calls cloud-based large models (e.g., Doubao, Wenxin Yiyan), returns results to the glasses, and plays them back. This approach significantly reduces on-device power consumption and cost.

Model Quantization Techniques

Regardless of platform, quantization is key to on-device deployment.

● INT8 and INT4: Compressing FP32 models into 8-bit or 4-bit integers. Testing on RK3588 shows that INT8 quantization can reduce model size by 74% while delivering multi-fold inference speedups with acceptable accuracy loss.

● Mixed Precision: NPUs from Qualcomm and Rockchip support mixed-precision computation, retaining higher precision (e.g., FP16) in critical layers and using lower precision (INT8) elsewhere to balance speed and accuracy.

Supply Chain Landscape and BOM Cost Analysis

The cost structure of AI smart glasses directly determines their market adoption rate.

BOM Cost Breakdown

Taking Ray-Ban Meta as an example, its estimated BOM cost is approximately USD 135–150. Qualcomm’s AR1 chip, memory, and RF components account for about 37%–50% of total cost, making them the single largest cost item. In contrast, domestic AI smart glasses adopting a “BES2800 + local ISP” solution can significantly reduce chip costs, keeping total BOM below RMB 1,000 (approximately USD 140) and achieving strong price competitiveness.

Table 2: Estimated Cost Structure of Typical AI Smart Glasses Solutions

Key Component Suppliers

● Memory: As larger models move on-device, higher storage speed and capacity are required. Vendors such as Longsys, Samsung, and Micron are introducing micro-packaged ePOP memory solutions tailored for wearables.

● Optics and Displays: While current AI glasses are mostly display-free, MicroLED and LCoS (potentially in next-generation Ray-Ban products) are gaining momentum. Companies like Goertek and Crystal-Optech play important roles in optical waveguide modules.

Future Technology Evolution and Outlook (2025–2030)

Neural Interfaces and Multimodal Interaction

Future AI glasses will move beyond voice and touch. Meta is exploring neural wristband interaction, using electromyography (EMG) signals for micro-gesture control. This places new demands on chips, requiring ultra-low-latency dedicated interfaces for high-frequency bioelectric signal processing.

Advanced Process Nodes and 3D Stacking

To further reduce power consumption, AI glasses chips will closely follow smartphone SoCs toward 3nm and even 2nm nodes. Meanwhile, 3D packaging technologies—such as stacking memory directly on top of SoCs—will become more common to reduce PCB area and data transfer power.

Privacy Computing

With cameras always on, privacy protection will become a legal red line. Future chips will integrate hardware-level privacy protection units, such as irreversible facial feature anonymization at the ISP level or physically unhackable camera indicator light circuits.

Conclusion

The AI smart glasses chip market in 2025 exhibits a clear three-tier pyramid structure:

● Apex (Flagship Experience): Dominated by Qualcomm Snapdragon AR1/AR2 and custom MediaTek chips, serving products from Meta, XREAL, and other leaders pursuing ultimate image quality, on-device large models, and future AR display capabilities. These solutions offer strong heterogeneous compute and mature ecosystems but come at higher cost.

● Middle Tier (Mass Adoption): Occupied by the BES2800 + domestic ISP (Sigmastar/SuperAcme) combination. With exceptional energy efficiency and cost advantages, this approach is accelerating the transition of AI glasses from tech novelties to mass consumer products and has become the mainstream choice for domestic brands.

● Base / Vertical Markets: Filled by Unisoc W517 (standalone connectivity), Rockchip RK3588 (heavy computation), and Realtek/Airoha (audio and entry-level vision). These platforms address niche demands such as standalone calling, industrial AR, and low-cost cross-border e-commerce products.

As on-device large-model technologies mature and chip energy efficiency continues to improve, smart glasses are approaching a moment reminiscent of the smartphone boom of 2007. The deep coupling between chip vendors, algorithm developers, and device brands will ultimately determine who defines the next generation of personal computing platforms.