DEEPX DX-M1 25-TOPS M.2 NPU ($139, 3.2W IQ8) & Milk-V Megrez 32GB RISC-V AI Motherboard ($199) Benchmark
Samsung 5nm DEEPX DX-M1 runs 16 concurrent 1080p YOLOv11 streams at 31.8 FPS on 3.2 Watts with 4GB onboard LPDDR5, while the $199 Milk-V Megrez brings 32GB LPDDR5 and a 20-TOPS NPU to a Mini-ITX SiFive P550 RISC-V motherboard.
⚡ Key Engineering & Supply-Chain Takeaways
- DEEPX DX-M1 ($139, M.2 2280 PCIe Gen3 x4) delivers 25 TOPS at 3.2W typical power (~7.8 TOPS/W) on Samsung 5nm silicon, hitting 248 FPS on YOLOv11s and co-hosting 4 vision/audio models in its 4GB onboard LPDDR5.
- DEEPX's IQ8 smart mixed-precision quantization preserves 99.6% of FP16 accuracy (-0.18% Top-1 drop) on Vision Transformers (ViT, CLIP, Depth-Anything-V2) without retraining.
- Milk-V Megrez ($199, Mini-ITX) pairs the ESWIN EIC7700X (4× out-of-order SiFive P550 RISC-V cores + 19.95 TOPS INT8/FP16 NPU) with up to 32GB LPDDR5, 32-ch H.265 decode, and a PCIe x8 GPU slot.
🔬 Technical & Supply-Chain Specification Matrix
| DEEPX DX-M1 M.2 NPU ($139) | Samsung 5nm / 25 TOPS @ 3.2W Typical / 4GB On-Module LPDDR5 / PCIe Gen3 x4 |
|---|---|
| DX-M1 Vision & ViT Efficiency | 248 FPS YOLOv11s (77.5 FPS/W) + IQ8 Quantization (-0.18% ViT Top-1 vs FP16) |
| Milk-V Megrez RISC-V ($199) | Mini-ITX ESWIN EIC7700X (4× SiFive P550 OoO @ 1.8GHz + 19.95 TOPS INT8/FP16 NPU) |
| Megrez Memory & Expansion | 16GB/32GB 128-Bit LPDDR5 + PCIe Gen3 x8 Slot + 32-Channel 1080p30 H.265 Decode |
5nm Sub-5W Vision NPUs and $199 32GB RISC-V AI Workstations
Official product specifications from DEEPX and Milk-V highlight two sub-$200 edge AI hardware platforms:
- DEEPX DX-M1 ($139, M.2 2280): Built on a Samsung 5nm process with 4GB of dedicated LPDDR5, the 25-TOPS DX-M1 sustains 248 FPS on YOLOv11s at just 3.2W while its IQ8 quantization engine runs
CLIP-ViT-B/16andDepth-Anything-V2within 0.18% of native FP16 accuracy. - Milk-V Megrez ($199, Mini-ITX): Powered by the ESWIN EIC7700X (
4× SiFive P550+19.95 TOPS NPU) with up to 32GB LPDDR5 and an open PCIe x8 slot, the Megrez gives developers a complete open-ISA RISC-V workstation for 7B–14B local LLMs and 32-channel video analytics.
🔗 Related AI Hardware in RECIWO Database
All 93 Devices →DEEPX DX-M1 25-TOPS Ultra-Low-Power M.2 Edge AI NPU
5nm M.2 2280 PCIe Gen3 x4 neural accelerator delivering 25 INT8 TOPS at just 3.0W–5.0W with 4GB onboard LPDDR5 and proprietary IQ8 quantization that matches FP16 vision/VLM accuracy for Raspberry Pi 5 and fanless industrial edge servers.
Milk-V Megrez Mini-ITX RISC-V AI PC Board (ESWIN EIC7700X)
Standard Mini-ITX open-architecture RISC-V AI workstation board powered by the ESWIN EIC7700X SoC (quad-core SiFive P550 out-of-order CPU + 19.95 TOPS INT8/FP16 NPU), supporting up to 32GB LPDDR5 and discrete PCIe x8 GPUs/accelerators.
Raspberry Pi AI HAT+ 2 (Hailo-10H 40 TOPS + 8GB RAM)
Second-generation official Raspberry Pi 5 GenAI HAT pairing the 40-TOPS (INT4) Hailo-10H neural processor with 8GB of dedicated onboard LPDDR4X memory—running 1.5B–3B local LLMs and VLMs at 3 Watts without touching host Pi RAM.
Radxa Fogwise AirBox Q900 (48 TOPS Edge AI Box)
Compact CNC-milled fanless aluminum micro-server powered by the 4nm Qualcomm QCS9075 / SG2300x series with 48 INT8 TOPS and dual 2.5GbE ports for home AI gateways.
📚 Related Deep-Dive Engineering Teardowns
Browse All 113 Articles →DEEPX DX-M1 (25 TOPS) vs Hailo-10H vs Milk-V Megrez RISC-V Benchmark: Sub-5W M.2 NPU & SiFive P550 Edge AI Tested
RECIWO Labs benchmarked the $139 DEEPX DX-M1 5nm M.2 NPU (25 TOPS at 3.2W, 4GB LPDDR5, IQ8 FP16-equivalent quantization) and the $199 Milk-V Megrez Mini-ITX RISC-V AI motherboard (ESWIN EIC7700X quad-core SiFive P550 + 19.95 TOPS NPU + 32GB LPDDR5) against the Raspberry Pi AI HAT+ 2 (Hailo-10H 40 TOPS).
Raspberry Pi AI HAT+ 2 (Hailo-10H 8GB) Review & Benchmark: Running 3B Local LLMs and Whisper in 3 Watts for $130
RECIWO Labs tested the official $130 Raspberry Pi AI HAT+ 2 powered by the 40-TOPS Hailo-10H GenAI NPU and 8GB of dedicated on-HAT LPDDR4X memory, benchmarking Llama-3.2-3B, Qwen2.5-Coder-3B, and Whisper latency against the first-gen Hailo-8 and Raspberry Pi 5 CPU inference.
RISC-V Vector (RVV 1.0) & Open-Standard Matrix Extensions in 2026 Edge AI Silicon
How SpacemiT K1/M1, Sophgo SG2380, and Kendryte K230 RISC-V SoCs use RVV 1.0 vector units and custom matrix tiles to run 1.5B–7B quantized models without Arm licensing fees.
