CN

M50 is a new-generation, energy-efficient edge AI chip designed for large-model inference. Built on Houmo.AI’s innovative compute-in-memory architecture, it enables local computing to better protect data privacy. A single M50 can smoothly run large models with up to 120 billion parameters, accelerating the deployment of large models across Agent Computers, smart mobile devices, embodied robots and edge inference systems, and enabling a wide range of edge devices to evolve into intelligent agents.

Specifications
Data Type
INT8 / INT16 / FP16
FP32 / bFP16 / bFP24
Physical Computing Power
160 TOPS @INT8
100 TFLOPS @bFP16
Typical Power Consumption
10W
Memory Capacity
12~48GB 192-bit LPDDR5/5X
Memory Bit Width
Max to 192-bit
Memory Bandwidth
Max to 153.6 GB/s
Host Channel
PCIe Gen4 X4
Inter-Chip Interconnection
HM-Link, 16GB/s
Power Management
Low Power MCU
Specifications
Data Type
INT8 / INT16 / FP16 / FP32 / bFP16 / bFP24
Physical Computing Power
160 TOPS @INT8, 100 TFLOPS @bFP16
Typical Power Consumption
10W
Memory Capacity
12~48GB 192-bit LPDDR5/5X
Memory Bit Width
Max to 192-bit
Memory Bandwidth
Max to 153.6 GB/s
Host Channel
PCIe Gen4 X4
Inter-Chip Interconnection
HM-Link, 16GB/s
Power Management
Low Power MCU