Innovation and Breakthrough of Homegrown AI Chips: Opportunities and Challenges in the Edge-Terminal Era

After the explosion of large AI models, compute is no longer confined to the cloud; more and more intelligent algorithms now run locally on edge devices. Smart cameras recognize human shapes and behaviors, in-vehicle terminals monitor driving in real time, industrial cameras auto-detect defects, and robot vacuums identify targets offline. Edge AI has become the fastest-growing, most widely deployed, and strongest home-replacement battlefield, giving domestic SoC vendors a window in multimedia and AI processing.

Strictly speaking, the edge-AI market splits into edge-terminal and edge-server segments. Edge servers are bulky, costly, and high-compute, serving smart parks and factory edge nodes. Edge terminals (this article’s focus) are mass-volume, cost-sensitive, and fragmented in scenarios. They sense the environment and process video/voice on the spot, delivering edge-AI functions and empowering hardware.

Which devices count as edge terminals? Smart cameras (IPC, smart doorbells, dash cams), industrial vision cameras and QC terminals, in-vehicle modules (AVM, DMS, DVR, ADAS assist), self-service retail kiosks, smart-home devices (speakers, vacuums, appliance controls), and smart-city edge nodes—all must run processing locally.

Edge-AI chips follow two paths: an SoC with built-in NPU for low-cost, low-power full intelligence, or a discrete AI accelerator that adds compute for multi-model, heavy-load professional inference.

SoC (System on Chip) integrates CPU, GPU, AI, video, audio, and peripherals in one die. Built-in NPU has become industry consensus and the most adopted edge-AI approach.

Rockchip RK3576, a general-purpose edge-AI SoC: 8-core CPU (4×A72 + 4×A53); 6-TOPS NPU (INT4/8/16, FP16); Mali-G52 GPU; 8-K decode, 4-K encode; multi MIPI-CSI for multi-camera, multi-display out (DSI, HDMI). Targets industrial tablets, AI cameras, vehicular DVR, robot vision.

HiSilicon Hi3403V100 couples AI-ISP (image enhancement + AI co-optimization). A pro-vision SoC with quad A55, 10-TOPS NPU tightly merged with ISP. High-spec ISP excels in back-light and low-light; multi 4-K video I/O; high deployment efficiency for detection/tracking.

How to split tasks efficiently among CPU, GPU, NPU and discrete accelerator—pre-processing on GPU/CPU, inference on NPU/accelerator, post-processing on CPU—is the key performance challenge. Hence AI accelerators were born, dedicated to inference and linked to the main SoC via PCIe. SoC handles system scheduling, video, graphics, UI; accelerator runs models and supplies AI compute.

Rockchip RK1820, an NPU coprocessor for high-performance edge AI, acts as the “second brain.” 20-TOPS NPU, standalone model execution, INT8/16/FP16; pairs with RK3576/3588 via PCIe for higher inference.

In edge AI, TOPS, CPU cores, and process node matter, but survival hinges on choosing the right track.

Rockchip’s aim is not the fastest chip but the richest ecosystem.
Full compute ladder: RV1103/1106 for light cameras; RV1126/1109 for security default; RK3576 for mid/high terminals; RK3588 for flagship edge; RK3688 for next-gen high-compute core. This matrix is the universal base for homegrown gear—from low-power IPC to industrial gateways, AR glasses, robots, education boxes, vehicular DMS/CMS.
Tech edge: balanced multimedia + ISP + AI; strong codecs, ISP, and mature RKNN tool-chain. Strategy is full-scene coverage, not single-point breakthrough.

Not for big models but for massive IoT and consumer devices.
Position: low-power, high-volume, cost-sensitive. Smart speakers (rich I2S/PDM/mic support), light edge-AI cameras, small appliance control, TTS/voice terminals. V853/V831: ultra-light AI SoC. R-series: control MCU-SoC. Allwinner chases “10-million-unit scenarios,” not TOPS.

Global leader in OTT boxes and smart-TV SoCs.
Position: home media hub + consumer smart device. AI is an enhancer; core strengths are video decode, HDR, A/V sync, TV/OTT ecosystem. Strong in smart projectors, conference bars, home entertainment AIOs.

Almost exclusively surveillance cameras.
Position: security-dedicated SoC. Strengths: strong ISP, strong compression, strict cost control, tight alignment with Hikvision/Dahua ecosystems. Flagships: FH8856, FH8852. Fullhan digs deep into the single huge surveillance赛道, winning on stability and cost.

MIPS-based, wearables and smart-home.
Position: feather-weight smart devices, tiny packages. Apps: smart doorbells, light IPC, kids’ watches, micro edge nodes. Traits: lowest power, high integration, small footprint. AISoC series for light vision inference.

For two years the talk was all TOPS—3, 6, 12—as if bigger numbers equal better chips. Core competence is never raw TOPS; it’s “interface matrix + scenario fit.”

In security cams, industrial cameras, smart doorbells, vehicular DMS/ADAS, what counts is: enough camera ports? (MIPI-CSI, DVP), how many video streams? real-time encode? (H.264/H.265/8K/4K), ISP tuning quality? In industrial DTU, smart gateway, robot, energy scenarios, peripherals trump TOPS: dual GbE/2.5G/RGMII/SGMII, RS232/485/CAN/UART, Wi-Fi/BT, 4G/5G modules, multiple USB/SPI/I2C. In smart control panels, aftermarket car displays, AR/VR, smart POS, priorities shift to display ports (MIPI-DSI, HDMI, eDP), multi-screen support, UI performance (GPU/graphics), with AI as helper, not star.

Aftermarket automotive keywords: shock resistance, voltage swing, ‑40-85 °C, eMMC lifetime, multi-CSI for DMS/OMS/ADAS, ms-level video latency.

Edge AI is the best track for homegrown chips; opportunity comes not from stacking TOPS but from knowing the scene and nailing the demand. Over the next few years, smart cameras, vehicles, industry, and home devices will be the stage where domestic chips prove themselves.