Report Measures DeepX M1 Edge Power Efficiency
Part 5 of the 'Edge AI Power' benchmarking series on Hackster.io, authored by Mario Bergeron, applies independent hardware power measurement to the DeepX DX-M1 edge AI inference chip. The series uses a repeatable INA228-based current-sensor methodology to measure real inference power draw across edge AI accelerators -- filling the gap for M.2 modules that lack built-in power telemetry. The DeepX DX-M1 is a South Korean NPU rated at 25 TOPS (INT8) with a 1-5W power envelope in M.2 2280 form factor over PCIe Gen3 x4. The series established its reference baseline with the Hailo-8 (Part 1), developed external INA228 measurement tools (Parts 2-3), measured Axelera Metis (Part 4), and now turns to the DX-M1. Part 6, covering the MemryX MX3, is already announced.
What this article covers
Part 5 of Mario Bergeron's 'Edge AI Power' benchmarking series on Hackster.io applies independent hardware power measurement to the DEEPX DX-M1 edge inference chip. The series' central problem: most M.2 AI accelerator modules lack the on-board power telemetry that Hailo provides natively, so there is no consistent cross-vendor way to compare actual inference power draw. Bergeron's solution is a hardware INA228-based current-sensor setup that intercepts the M.2 PCIe power rail, paired with his open-source mb-powermon.py utility (Apache-2.0, hosted on GitHub).
Series structure
The six-part series progresses as follows: Part 1 used the Hailo-8 to establish the reference methodology, exploiting its native on-board telemetry. Parts 2-3 built and validated an external measurement approach using an ElmorLabs power-insertion board and an INA228 current sensor, enabling the same rigor for modules without built-in telemetry. Part 4 applied that methodology to the Axelera Metis. Part 5 (this article) targets the DEEPX DX-M1. Part 6, announced at the end of Part 1, will cover the MemryX MX3.
DX-M1 chip specifications
The DX-M1 is a mass-production edge AI NPU from South Korean semiconductor startup DEEPX. Per vendor specifications: 25 TOPS at INT8 precision, 1-5W power draw (1W minimum at idle, 5W peak under load), M.2 2280 (M Key) form factor at 22x80mm, PCIe Gen3 x4 host interface (backward-compatible with Gen 1/2 and x1/x2 lane widths), up to 8GB LPDDR4x or LPDDR5 memory (the standard M.2 module ships with 4GB LPDDR5 at 5600 MT/s), and OS support for Ubuntu 24.04/22.04/20.04 LTS, Windows 11/10, Yocto, and Docker on both x86 and ARM host architectures. Supported AI frameworks include Ultralytics, TensorFlow, PyTorch, ONNX, and Keras. DEEPX claims roughly 20x power-performance efficiency versus GPGPUs -- a vendor-reported figure.
Why this benchmark matters
Vendor TOPS ratings and TDP ranges do not capture actual inference power at production batch sizes and operating temperatures. Edge deployments -- cameras, industrial controllers, robotics, mobile platforms -- carry hard power budgets; a chip that draws 4W instead of 2W under sustained load can fail the thermal and battery design envelope. Independent community measurement on a consistent testbed (same PCIe slot, same INA228 probe, same host x86 platform, same inference workload) gives practitioners a direct, apples-to-apples figure. The DX-M1 competes in the crowded M.2 edge inference market against the Hailo-8, Axelera Metis, and MemryX MX3 -- all of which appear or are scheduled to appear in this same benchmarking series.
Scoring Rationale
Independent community hardware benchmark for a niche but growing category: M.2 edge AI accelerators. Useful to practitioners selecting edge inference hardware on real power budgets, but limited in scope to one device in a six-part series. The ongoing Hackster series provides practitioner-relevant measurement data beyond vendor TDP specs; scores in the solid-but-niche range.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
