DEEPX DX-M1 Edge AI Accelerator and DXNN SDK

DEEPX is a fabless edge AI chip company headquartered in Seongnam, Gyeonggi-do, South Korea, led by founder and CEO Lokwon Kim, who previously designed semiconductors at Apple, IBM and Cisco. It builds low-power neural processing units so cameras, robots, drones and industrial PCs can run AI locally instead of relying on power-hungry GPUs or the cloud.

The DX-M1 chip delivers 25 TOPS (INT8) at roughly 1-5 W and is sold as a chip or as the DX-M1M M.2 2280 module with 4 GB LPDDR5 over PCIe Gen3 x4, working with x86 and Arm hosts on Windows, Ubuntu, Yocto and Android. The DXNN SDK compiles models from PyTorch, TensorFlow, ONNX, Keras and Ultralytics, and the company's IQ8 INT8 quantization is designed to keep accuracy loss under 1%. The line-up also includes the DX-H1 for edge servers and the DX-M2 under development for generative AI on battery-powered devices. Device makers such as Advantech, AAEON, ASRock and MSI appear on its partner list.

• DX-M1 NPU: 25 TOPS INT8 at about 1-5 W with up to 8 GB LPDDR4x or LPDDR5 memory.
• M.2 module: DX-M1M in M.2 2280 form factor with PCIe Gen3 x4 for industrial PCs and edge servers.
• DXNN SDK: compiler and runtime that import PyTorch, TensorFlow, ONNX, Keras and Ultralytics models.
• IQ8 quantization: INT8 technique designed to keep accuracy loss below 1% versus FP32.
• Broad OS support: Windows, Ubuntu, Yocto Linux, Android and Docker on x86 or Arm hosts.
• On-device language models: DX-M1 runs small language models and multi-channel vision-language processing locally.

A Vietnamese electronics or garment factory can add DX-M1 modules to existing industrial PCs for visual defect inspection on production lines. A camera or NVR maker can embed the chip to offer multi-channel intrusion and traffic analytics without a GPU. A robotics or drone integrator can run detection and small language models on board for offline operation.

Runs object detection and vision-language workloads at about 5 W, which DEEPX says is around 20 times more efficient than GPGPUs, cutting power and cooling needs in camera boxes and industrial PCs. The M.2 module can offload 4-channel detection on Intel Atom-class hosts at under 25% CPU, extending the life of low-cost hardware. Industrial temperature grades from -40 to 85 degrees Celsius suit factories and outdoor cabinets. Support for common frameworks shortens porting of existing models.

Physical AI in robots, drones and smart cameras is pushing inference out of data centres and onto devices with tight power budgets. Asian device makers and system integrators are looking for alternatives to GPU modules that are costly and power-hungry, and Southeast Asian factories and cities expanding camera networks need low-power accelerators that fit existing industrial PCs.

Headquartered in Seongnam, South Korea. Reports over 400 patents filed and more than 130 granted. Entered 5 nm mass production in 2025 and signed distribution agreements in the US, China, Taiwan and Europe, including DigiKey. Awards include CES Innovation Awards 2026 (double honoree) and 2024 (triple honoree), and a Korean presidential citation from KIPO. Partners include Samsung Foundry, Rambus, AAEON and a strategic partner in Japan.