Models & Pricing
DX-M1 — NPU Chip
- 25 TOPS (INT8)
- PCIe Gen3 x4 host interface
- Up to 8GB LPDDR4x @ 4266 MT/s or LPDDR5 @ 6000 MT/s
- 1W min – 5W max power
- FC-BGA 17 × 17 mm, 625-ball
- Commercial, industrial and AEC-Q100 Grade 3
DX-M1 M.2 2280 Module
- 25 TOPS (INT8)
- 4GB LPDDR5 @ 5600 MT/s
- PCIe Gen3 x4
- M.2 2280 M-Key, 22 × 80 mm
- 1–5W, –40 to 85°C
- Optional heatsink
DX-M1M M.2 2242 Module
- 25 TOPS (INT8)
- 512MB–1GB LPDDR4x integrated
- PCIe Gen3 x2
- M.2 2242 M+B Key
- 3W typical, –40 to 85°C
- Fastest integration for compact designs
Specifications
NPU
DEEPX DX-M series edge AI processor
AI Performance
25 TOPS (INT8)
Host Interface
PCIe Gen3 x4 (DX-M1) / PCIe Gen3 x2 (DX-M1M)
Memory
4GB LPDDR5 5600 MT/s (DX-M1 M.2), 512MB–1GB LPDDR4x (DX-M1M)
Power Consumption
1W min – 5W max (chip); 3W typical (DX-M1M)
Operating Temperature
–40 to 85°C industrial; AEC-Q100 Grade 3 standard
Form Factor
M.2 2280 M-Key (DX-M1 M.2); M.2 2242 M+B Key (DX-M1M); FC-BGA chip
Efficiency
~20× power-performance efficiency vs GPGPU
Accuracy
Under 1% accuracy drop versus FP32 at INT8
OS Support
Windows 11/10; Ubuntu 24.04/22.04/20.04 LTS; Yocto; Docker
AI Frameworks
Ultralytics, TensorFlow, PyTorch, ONNX, Keras
Host Support
x86 and Arm based architectures
SDK
DX-Compiler v2.0.0 and DX-RT v3.0.0 runtime
Thermal Design
Holds performance in industrial temperature range where competing parts throttle
Module Partners
Radxa AICore DX-M1M, DFRobot DX-M1
Target Hosts
Industrial PCs, edge gateways, SBCs, cameras, robotics controllers
Positioning
Cloud-class inference at 1–5W for always-on edge devices
Availability
Sampling and volume production through authorised distributors
Overview
DEEPX is a Korean edge-AI semiconductor company whose DX-M series targets a specific gap: edge devices that need real neural-network throughput but have no power or thermal budget for a GPU. The DX-M1 delivers 25 TOPS of INT8 inference while drawing as little as 1W and no more than 5W, which the vendor characterises as roughly 20 times the power-performance efficiency of a comparable GPGPU.
Two deployment shapes are available. The DX-M1 chip in a 17 × 17 mm FC-BGA package is for customers building their own boards and carrier designs, with up to 8GB of LPDDR4x or LPDDR5 attached. The DX-M1 M.2 2280 module packages the same 25 TOPS with 4GB of LPDDR5 behind a PCIe Gen3 x4 interface, so it can be dropped into an industrial PC, gateway or single-board computer without custom hardware. A shorter DX-M1M variant in M.2 2242 form factor uses PCIe Gen3 x2, integrated LPDDR4x and only 3W typical for space- and power-constrained designs — Radxa sells this as the AICore DX-M1M.
The industrial story is the differentiator. The DX-M1 is offered in commercial, industrial (–40 to 85°C) and AEC-Q100 Grade 3 automotive grades, with Grade 2 available on request. DEEPX claims competing NPUs throttle or fail in these temperature ranges while the DX-M1 sustains performance, which matters for factory-floor, in-vehicle and outdoor camera deployments.
Software support is unusually broad for an NPU vendor: Windows 10/11 as well as Ubuntu LTS and Yocto, with Ultralytics, TensorFlow, PyTorch, ONNX and Keras front-ends feeding the DX-Compiler and DX-RT runtime. Both x86 and Arm hosts are supported, so the module works equally well behind an Intel industrial PC or an Arm SBC. DEEPX also ships AI gateways, SBCs, SoM boards and industrial PCs built around the same silicon for customers who prefer a finished system.
Target Applications
Edge cameras and video analytics, smart factory quality control and defect detection, smart city traffic and crowd monitoring, robotics and AMR perception, drones, retail analytics, smart home video, and any x86 or Arm edge platform that needs 25 TOPS of INT8 inference inside a 5W thermal envelope.
Related Products
Request a Quote — DEEPX DX-M1 M.2
QS Compute — global B2B sourcing for DEEPX DX-M1 / DX-M1M NPU modules, evaluation units and volume allocation. Industrial and automotive grades available on request.
Get Your Quote →