DX-M1
Vision and language AI, on one chip.
4B-class intelligence at the edge.
Overview
DX-M1 brings computer vision, language generation, and visual understanding to edge devices. Delivering 25 TOPS of INT8 AI compute, it combines a compact design with dedicated AI memory to support a wider range of on-device applications.
Building on its vision capabilities, DX-M1 now runs small language models and vision-language models up to 4B parameters. DEEPX has measured 17 language and vision-language models on x86 and Raspberry Pi 5 reference systems, Available as a chip or M.2 module, it offers integration options for industrial PCs, robotics, kiosks, drones, and other edge systems.
From image-based questions to text generation and summaries, DX-M1 gives developers more ways to build intelligence into their products. real-time intelligence becomes universal, affordable, and inevitable.
Features | Details |
|---|---|
AI Performance | 25 TOPS (INT8) |
Host Interface | PCIe Gen3 x4 (Supports Gen 1/2/3 & x1/x2/x4) |
Memory Interface | LPDDR4x or LPDDR5 |
Memory Capacity : Up to 8GB | |
Operating Temperature | 0 ~ 70°C (Commercial) |
-40 ~ 85°C (Industrial) | |
-40 ~ 85°C (AEC-Q100 Grade3) | |
Supports AEC-Q100 Grade upon request | |
Package | FC-BGA, 17 x 17mm 625-ball, 0.65mm Pitch |
OS Support | Windows 11/10 |
Debian-based Linux (Ubuntu 24.04/22.04/20.04 LTS), Docker | |
Yocto Linux | |
Android | |
AI Frameworks | Ultralytics, PyTorch, TensorFlow, ONNX, Keras |
System Support | x86 and ARM-based Architectures |
Key Benefits
Expanded AI CapabilitiesExtend edge applications beyond computer vision with language generation and visual understanding. DEEPX has measured 17 language and vision-language models on DX-M1, expanding model options for text responses, image-based questions, and summaries.
Efficient 4B-Class InferenceRun selected 4B-class language and vision-language models at over 10 tokens per second on DX-M1. Optimized inference brings text generation and visual understanding to compact edge systems, with performance measured across two host platforms.
Flexible Host IntegrationIntegrate DX-M1 with x86 or ARM-based systems through PCIe. Language model benchmarks on x86 PCs and Raspberry Pi 5 demonstrate deployment options across different host platforms, helping developers evaluate configurations for their applications.
Compact, Low-Power DesignBuild compact edge systems with a DX-M1 M.2 module rated at 5W maximum. Dedicated LPDDR5 memory keeps AI model storage separate from host memory, supporting integration into industrial PCs and other space-constrained devices.
Performance
- DEEPX DX-M1 (INT8, Batch=1)
- Company A (INT8, Batch=1)
- Nvidia Xavier (INT8, Batch=1)
- This graph shows the benchmark results for Batch=1, all tested under the same conditions.
- Xavier measurements were conducted at room temperature on an NVIDIA Jetson AGX Xavier development kit. The benchmarks utilized INT8 precision with the "no batch" (default=1) option, as recommended by NVIDIA.
- Company A results used the optimize option of "random-calib-set" with default values for the parser and compiler.
- DEEPX DX-M1 numbers are based on the SDK (September 2025), DX-Compiler v.2.0.0, and DX-RT v.3.0.0.
- Company A and DX-M1 were measured at room temperature on a single device through a PCIe interface on an x86 Host PC (Intel® Core™ i7-14700K CPU @ 5.6GHz).
Related Products
Discover scalable AI solutions — chips, modules, and systems — that integrate seamlessly with DX-M1.
DX-M Series
AI Chips
DX-M Series
AI Modules
DX-M Series
Single Board Computers
DX-M Series
SoM Boards
DX-M Series
AI Gateway
DX-M Series
Industrial PCs
Use Cases
Edge Cameras Systems
Real-time surveillance and intrusion detection without cloud latency, ensuring immediate threat response and local privacy protection
Smart Mobility
Autonomous navigation and collision avoidance without network dependency, ensuring reliable operation regardless of connectivity conditions
Smart Factory
On-site quality control and defect detection with superior thermal efficiency, operating reliably in harsh industrial environments without cooling systems
Smart Cities
Live traffic management and crowd monitoring with reduced bandwidth costs, processing thousands of camera feeds locally
Robotics
Extended battery life and compact form factor enable longer autonomous missions with complex AI processing for mobile robots
Drones
Lightweight, power-efficient AI processing for aerial surveillance and autonomous flight control without draining battery life
Edge Computing
High-performance AI inference at network edge with minimal power consumption, reducing server loads and network bandwidth requirements
Smart Homes
Local video analytics and scene recognition without sending personal data to cloud servers, ensuring family privacy
Smart Retail
Real-time customer analytics and inventory management through local video processing, protecting customer privacy while optimizing operations
Trusted by Global Leaders
Global partners collaborate with DEEPX to bring on-device AI into real-world applications worldwide