Zhongda Tech K3 Industrial Edge AI Module: 60 TOPS RISC-V on Mini COM-E
On August 27, 2026, Zhongda Tech announced an upcoming industrial edge AI compute module built on the SpacemiT K3 RISC-V SoC. In a standard Mini COM-E form factor (84×55 mm), it delivers 130 KDMIPS of general-purpose compute and 60 TOPS of AI inference — enough to run a 30-billion-parameter LLM locally. With a 100% domestic supply chain, -40°C to 70°C industrial temperature range, and rich industrial I/O including 8× CAN-FD, this module brings RISC-V AI to the factory floor.
Overview: RISC-V AI at the Edge
Zhongda Tech (众达科技), a Chinese embedded systems specialist, partnered with SpacemiT (进迮时空) to build a fully domestic industrial edge AI module based on the K3 SoC. The module targets applications where data must stay on-premises — smart manufacturing, humanoid robotics, power-grid monitoring, and edge AI gateways — combining general-purpose RISC-V computing with on-die AI acceleration in a single chip.
The K3 is the world's first mass-produced RISC-V SoC to conform to the RVA23 profile, integrating 8 high-performance X100 CPU cores alongside 8 A100 AI cores. This "homogeneous fusion" architecture means both the host CPU and the AI accelerator speak the same RISC-V ISA, simplifying the software stack and eliminating cross-architecture data-movement overhead.
Full Specifications
| Module Specifications | |
|---|---|
| SoC | SpacemiT K3 (RISC-V, RVA23 Profile) |
| CPU Cores | 8× X100 high-performance cores, up to 2.2 GHz |
| CPU Performance | 130 KDMIPS (general-purpose) |
| AI Cores | 8× A100 ultra-wide parallel RISC-V AI cores |
| AI Performance | 60 TOPS (general-purpose AI compute) |
| Vector Width | 1024-bit RVV 1.0 (RISC-V Vector extension) |
| Supported Data Types | BF16 / FP16 / FP8 / INT8 / INT4 |
| Max LLM Size | 30B parameters (300 billion), runs locally |
| Graphics | Integrated 3D graphics engine |
| Video Codec | H.265 / H.264 decode, up to 4K@180fps |
| Memory | 32 GB LPDDR5 (domestic brand, on-board) |
| Crypto Engine | International algorithms + Chinese commercial cryptography (SM2/SM3/SM4) |
| Form Factor | Mini COM-E (84 mm × 55 mm) |
| Power Consumption | 13 W (typical) |
| Temperature Range | -40°C to +70°C (industrial grade) |
| Input Voltage | 4.5 V to 16 V (wide-range DC input) |
| Supply Chain | 100% domestic components, all-surface-mount design |
Industrial I/O Interfaces
The module's I/O is tailored for industrial and robotics applications, with heavy emphasis on real-time control buses:
| Interface | Count / Detail |
|---|---|
| CAN-FD / CAN 2.0B | 8 channels (independent) |
| TTL Serial (UART) | 4 channels |
| I²C | 6 channels |
| Gigabit Ethernet | 2 ports (optical-electrical combo) |
| PCIe | PCIe 3.0 ×8 |
| USB | 4× USB 2.0 (multiplexed with USB 3.0) |
| Type-C | Full-function (data + display + power) |
| Camera Input | 2× 4-lane MIPI-CSI |
| Display Output | Multiple display outputs |
Eight CAN-FD channels are particularly significant for multi-axis robotics — a single module can independently control up to 8 servo drives or joint motors without an external CAN expander, a configuration that is uncommon even among ARM-based industrial SoMs.
AI Software Stack
The module ships with a complete AI SDK designed for rapid deployment:
- Native llama.cpp support — GGUF/GGML model format, verified running Qwen3-30B locally with smooth inference throughput
- Dual-model parallel inference — run two models simultaneously (e.g., an LLM + a vision model) on the same chip
- ONNXRuntime — built-in inference engine for standard ONNX models
- YOLO series — real-time object detection out of the box
- VLM (Vision-Language Models) — multimodal inference for image+text tasks
- ASR / TTS — speech recognition and synthesis for voice interaction
- RAG local knowledge base — deploy retrieval-augmented generation entirely on-device
Operating system support includes standard Linux and Kylin V10 (银河), the Chinese government-certified domestic OS, making the module suitable for Xinchuang (信创) procurement scenarios that mandate domestic software stacks.
Running a 30B Model on the Edge
The headline capability is local 30B-parameter LLM inference. The K3's 8 A100 AI cores support FP8 and INT4 quantization natively — a 30B model in INT4 requires approximately 15 GB of memory, which fits comfortably within the 32 GB LPDDR5 pool. The 1024-bit RVV 1.0 vector width enables wide SIMD-style parallelism across all 8 AI cores simultaneously.
For context, the K3's predecessor K1 (8× X60 cores, 2 TOPS AI) could only run small models (<1B). The K3 represents a 30× improvement in AI compute and an 80× increase in supported model size over the K1 — a generational leap that makes edge-deployed LLMs practical for the first time on RISC-V hardware.
Comparison: K1 vs. K3 vs. Module
| Parameter | SpacemiT K1 | SpacemiT K3 | Zhongda K3 Module |
|---|---|---|---|
| CPU Cores | 8× X60 @ 1.6 GHz | 8× X100 @ 2.2 GHz | 8× X100 @ 2.2 GHz |
| CPU IPC (SpecInt2k6/GHz) | ~4 | ~9 | ~9 |
| AI Cores | — | 8× A100 | 8× A100 |
| AI Compute | 2 TOPS | 60 TOPS | 60 TOPS |
| RVV Width | 256-bit | 1024-bit | 1024-bit |
| Max LLM | <1B | 30B | 30B (Qwen3-30B verified) |
| FP8 Inference | No | Yes (native) | Yes (native) |
| RVA23 Profile | No | Yes (first mass-produced) | Yes |
| Form Factor | SoC only | SoC only | Mini COM-E 84×55 mm |
| Temperature | Commercial | Commercial | -40 to +70°C industrial |
| Power | — | — | 13 W typical |
| CAN-FD | — | — | 8 channels |
Application Scenarios
- AI Edge Computing Gateway / Industrial Server — local inference with data staying on-premises; low latency, enhanced security
- Embodied Intelligence / Humanoid Robots — edge perception and decision-making; 8× CAN-FD directly drives joint motors
- AI Computer / Personal Computing Server — on-device large models for individual or enterprise AI compute
- Power Grid, Smart Manufacturing — multi-bus connectivity (CAN/serial/Ethernet) for device interconnection and predictive maintenance
Domestic Supply Chain Significance
The module is designed for China's Xinchuang (Information Technology Application Innovation) initiative, which mandates fully domestic supply chains for government and critical-infrastructure procurement. Key differentiators:
- 100% domestic components — from the K3 SoC down to passive components, all sourced from Chinese brands
- Domestic LPDDR5 memory — 32 GB on-board, avoiding foreign DRAM supply risk
- Hardware crypto engine — supports both international algorithms (AES, RSA, SHA) and Chinese commercial cryptography standards (SM2, SM3, SM4) for compliant secure boot and data-at-rest encryption
- All-surface-mount design — no BGA rework dependencies on foreign assembly equipment
SpacemiT K3 SoC Architecture Context
The K3 SoC at the heart of this module was announced in January 2026 as the world's first mass-produced RISC-V chip conforming to the RVA23 profile. Its key architectural milestones:
- First RVA23-conformant mass-produced chip — the RISC-V International profile ratified in October 2024 that defines a stable software target for OS and application developers
- First 1024-bit RVV RISC-V chip — the widest vector engine in any shipping RISC-V silicon
- First FP8 native inference — hardware support for 8-bit floating-point, the emerging standard for efficient LLM inference
- First full chip-level virtualization — hardware virtualization extensions for multi-tenant edge deployments
SpacemiT's CPU core roadmap extends beyond K3: the X200 core (development complete, targeting 2027 mass production) reaches 16 SpecInt2k6/GHz — comparable to ARM's Neoverse N2 server core — with the next-generation X300 targeting 20/GHz. The AI core roadmap scales from A60 (2 TOPS, edge) → A100 (32 TOPS, desktop) → A200 (256 TOPS, server). On the SoC side, K3 is followed by K5 and K7 (edge), while the V-series (V100 taped out, V200 in development) targets datacenter servers.
Official Resources
- SpacemiT K3 Pico-ITX Development Kit — official reference board
- SpacemiT Documentation Center — K3 product brief, specs, and SDK docs
- SpacemiT Developer Forum — community discussions and demos
- Zhongda Tech announcement (CSDN, Aug 27, 2026) — original source for this article