MemryX MX3 M.2 (24 TFLOPS DRAM-Free At-Memory Dataflow) & Kinara Ara-2 / NXP Ara-240 (40 TOPS, 16GB LPDDR4X) NPU Comparison
Silicon architecture comparison of the $149 MemryX MX3 M.2 2280 card (4× cascaded MX3 chips delivering 24 BF16/INT8 TFLOPS at ~8W with zero external DRAM) and the $188 Kinara Ara-2 / NXP Ara-240 M.2 module (40 TOPS polymorphic GenAI NPU with 16GB LPDDR4X over PCIe Gen4 x4).
⚡ Key Engineering & Supply-Chain Takeaways
- MemryX MX3 M.2 (MX3-2280-M-4, $149) cascades four MX3 chips to deliver 24 TFLOPS/TOPS of mixed BF16/INT8 compute at ~8W typical card power (~2W/chip) with zero external DRAM.
- By storing 42M–80M model weights directly in on-die SRAM across the 4-chip spatial dataflow pipeline, MemryX eliminates off-chip memory bottlenecks and compiles 1,000+ ONNX/PyTorch models without retraining.
- Kinara Ara-2 / NXP Ara-240 ($188, M.2 2280 PCIe Gen4 x4) pairs a 40-TOPS polymorphic NPU with up to 16GB of on-module LPDDR4X DRAM to run local 7B LLMs (Llama-3) and multimodal VLMs (LLaVA) within 3W–12W.
🔬 Technical & Supply-Chain Specification Matrix
| MemryX MX3 M.2 ($149) | 4× Cascaded MX3 Chips / 24 TFLOPS BF16+INT8 / ~8.0W Typical (~2.0W/Chip) |
|---|---|
| MX3 Memory & Compiler | Zero External DRAM (100% On-Die SRAM, 42M–80M Params) / Zero-Retraining SDK |
| Kinara Ara-2 / NXP Ara-240 ($188) | 40 TOPS Polymorphic GenAI NPU / PCIe Gen4 x4 / 3W Typical (8W–12W LLM TDP) |
| Ara-2 Memory & Workloads | Up to 16GB On-Module LPDDR4X / Local 7B Llama-3, Phi-3, LLaVA & ViT Execution |
DRAM-Free At-Memory Spatial Dataflow vs 16GB LPDDR4X Edge GenAI NPUs
Datasheets from MemryX and Kinara / NXP contrast two M.2 2280 acceleration strategies:
- MemryX MX3 M.2 ($149): Stores 42M–80M weights in 100% on-die SRAM across 4 cascaded chips (24 TFLOPS BF16/INT8 at ~8W, zero external DRAM).
- Kinara Ara-2 / NXP Ara-240 ($188): Pairs a 40-TOPS polymorphic NPU with 16GB of on-card LPDDR4X over PCIe Gen4 x4 to run local 7B LLMs and VLMs.
🔗 Related AI Hardware in RECIWO Database
All 93 Devices →MemryX MX3 M.2 2280 (MX3-2280-M-4) 24-TFLOPS At-Memory AI Card
M.2 2280 M-Key edge AI accelerator module cascading four MemryX MX3 At-Memory Dataflow chips to deliver 24 TFLOPS/TOPS of mixed BF16/INT8 compute at ~8W typical power (~2W per chip) with zero external DRAM required and push-button ONNX/PyTorch compilation.
Kinara Ara-2 (NXP Ara-240) 40-TOPS 16GB Edge GenAI M.2 Module
Discrete Edge GenAI M.2 2280 PCIe Gen4 x4 accelerator powered by the Kinara Ara-2 (NXP Ara-240) polymorphic dataflow NPU, pairing 40 INT8/FP16 TOPS with up to 16GB of on-module LPDDR4X DRAM to run local 7B LLMs, Vision Transformers, and multimodal VLMs at 3W–10W.
Axelera Metis M.2 Max 214-TOPS D-IMC Edge AI Accelerator
M.2 2280 PCIe Gen3 x4 edge AI accelerator powered by Axelera's Metis AIPU (Quad-Core RISC-V + Digital In-Memory Computing), delivering up to 214 INT8 TOPS at 15 TOPS/W (~6.5W typical) for multi-stream computer vision and local VLM/LLM inference.
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.
📚 Related Deep-Dive Engineering Teardowns
Browse All 113 Articles →MemryX MX3 vs Kinara Ara-2 (NXP) vs Axelera Metis M.2 Comparison: DRAM-Free At-Memory Dataflow vs 16GB 40-TOPS Edge GenAI NPU
A technical silicon architecture comparison of the MemryX MX3 M.2 2280 ($149, 4-chip 24-TFLOPS BF16/INT8 At-Memory Dataflow accelerator requiring zero external DRAM at ~8W), Kinara Ara-2 / NXP Ara-240 M.2 ($188, 40-TOPS polymorphic Edge GenAI NPU with 16GB LPDDR4X for local LLMs and VLMs), and Axelera Metis M.2 Max ($189).
Axelera Metis M.2 Max vs SiMa.ai Modalix vs DEEPX DX-M1 Comparison: 214-TOPS D-IMC vs 6nm Multimodal Physical AI SoM
A technical architecture and specification comparison of the Axelera Metis M.2 Max ($189, 214 INT8 TOPS Digital In-Memory Computing AIPU at 15 TOPS/W), SiMa.ai Modalix DevKit 3.0 ($999, TSMC 6nm 50-TOPS standalone Physical AI SoM with 4x MIPI CSI-2), DEEPX DX-M1 ($139, 25 TOPS at 3.2W), and Hailo-10H ($130).
LPDDR5X vs. LPDDR6 & Processing-in-Memory (PIM) for Edge AI Devices: Breaking the Von Neumann Wall
How on-DRAM vector multiply-accumulate units in 2026 LPDDR5X-PIM and LPDDR6 memory cut picojoules-per-bit by 54% during mobile and robotics LLM decoding.
