AX650N-Intelligent Edge AI Box
developed by YIYUE Tech, featuring Axera’s third-generation high-efficiency AI vision
chip, the AX650N. This chip delivers a powerful AI computing performance of 72 TOPS
@ INT4 or 18 TOPS @ INT8, capable of running popular large-scale models. In addition
to efficiently supporting CNN networks, it natively supports various large-scale model
architectures such as Transformer and Conformer.
box achieves the following results with common open-source AI models:
•
YOLOv5s: 130 FPS (7.66 ms) with 640x640 resolution input.
•
MobileNetV2: Over 1798 FPS (0.556 ms) with 224x224 resolution input.
The built-in ISP supports 8K@30fps, along with Axera’s exclusive AI ISP for low-light
level night vision capabilities. It also includes H.264 and H.265 codecs, supporting up to
8K@60fps decoding and 8K@30fps encoding. Real-world testing demonstrates stable
performance with 32 channels of 1080P@30fps decoding and 16 channels of
1080P@30fps encoding.
In terms of interfaces, the AX650N supports 64-bit LPDDR4x, multiple MIPI inputs,
Gigabit Ethernet, USB 3.0, and HDMI 2.0b output. With its built-in high computing
power and robust codec capabilities, it meets the industry’s demand for high-
performance edge AI computing.
Equipped with various deep learning algorithms, the box enables applications such as
visual structuring, behavior analysis, and state detection. It efficiently supports
Transformer models, large vision models, and DeepSeek models. An SDK is provided,
allowing users to perform secondary development and quickly achieve productization.
System Resources | Parameter Description |
CPU | Octa-core Cortex-A55 1.7GHz, supports NEON acceleration, integrated FPU processing unit |
NPU | Supports up to 72 TOPS @ INT4 or 18 TOPS @ INT8; |
CODEC | Supports H.264/H.265 codec; |
ISP | Supports up to 8192x4320@30fps; |
DSP | Dual-core 800MHz |
RAM | 8GB LPDDR4 (Customizable based on user requirements) |
ROM | 8GB eMMC (customizable based on user requirements) |
Operating System | Linux (Supports Ubuntu/Debian desktop systems) |
Interface Resources | Parameter Description |
Network Interface | Gigabit RJ45 ports *2 |
USB Interface | USB 3.0 *1, USB 2.0 *1 |
Serial Interface | RS232 *1, RS485 *1 |
Alarm Interface | 2 channel switch input, 2 channel switch output |
Audio Interface | Audio input *1, Audio output *1 |
PCIe Interface | Built-in PCIe slot *1 (supports 4G communication module) |
SATA Interface | Built-in M.2 SATA slot *1 |
MIPI Interface | MIPI-CSI camera interfaces 2 * 4-lane |
SIM Slot | Built-in SIM card slot *1 |
eMMC Interface | Built-in TF card slot *1 |
Type-C Interface | Built-in Type-C Debug interface *1 |
Video Interface | HDMI v2.0 output interface *1 |
Power Interface | DC 12V/2A power input interface *1 |
Reset Port | Built-in reset port *1 |
Physical Dimension | Parameter Description |
Dimensions (L×W×H) | 180mm* 148mm* 37mm |
Weight | 800g |
utilizing only 80% of NPU resources even under 32-channel scenarios.
The NPU demonstrates significant performance advantages across various
models.
Extremely low power consumption, with a typical value of 5W.
Under high-temperature conditions (85°C), the 32-channel AI BOX demo runs
stably at full load, with the CPU core temperature remaining below 100°C.
Contact us