Grįžti į MemryX

MemryX MX3-2280-M-x-I MX3 M.2 AI Accelerator Modules

MemryX MX3-2280-M-x-I MX3 M.2 AI Accelerator Modules are powerful, energy-efficient solutions designed to facilitate high-performance Artificial Intelligence (AI) inference for edge servers. Engineered as companion modules, the MemryX MX3-2280-M-x-I modules significantly alleviate the processing load of deep neural network (DNN) computer vision (CV) models from a host CPU. A distinctive dataflow architecture excels in real-time, low-latency inferencing, saving system power.

The MX3 M.2 2280 modules are compatible with the industry-compliant M.2 form factor and can be plugged into existing systems with an M.2 socket with no hardware changes. Each M.2 module contains four MemryX MX3 Edge AI Accelerators to execute inferencing in real-time for a broad range of applications and market segments. Designed to meet the demands of modern AI applications, the MemryX M.2 AI accelerator modules set performance, power efficiency, and adaptability benchmarks.

Features

  • Supports all common frameworks
  • Dataflow architecture for ultra-low latency
  • Advanced power management
  • Up to 80 million weight parameters
  • On-chip storage of model parameters and matrix operators eliminates the need for external DRAM
  • Two 4-lane PCIe Gen3 for up to 4GB/s bandwidth
  • Supports multiple concurrent models
  • Floating-point activations for high accuracy
  • End-user models can be deployed as-is without quantization, pruning, compression, or retraining
  • Options
    • MX3-2280-M-4: 4-chip M.2 module, 22mm x 80mm, M-Key
    • MX3-2280-M-2: 2-chip M.2 module, 22mm x 80mm, M-Key

Applications

  • Industrial 4.0 and robotics
  • Automotive
  • Internet of Things (IoT)
  • Metaverse
  • Smart vision systems
  • Computing devices

Specifications

  • NGFF M.2 2280 M Key Socket 3 form factor
  • 3.15” x 0.87” (22mm x 80mm) dimensions
  • 3.3V ±5% input voltage
  • 8W typical power, 14W maximum (M.2 limit)
  • Interface
    • Two 2-lane PCIe Gen 3, 8GT/s per PCIe lane
    • Two USB Gen 3
  • Four MX3s ICs
  • ONNX, PyTorch, TensorFlow, Keras, and TensorFlow Lite frameworks
  • Weight parameters
    • 80 million, 4-bit
    • 40 million, 8-bit
  • Any host processor including ARM, x86, and RISC-V
  • OS support
    • Ubuntu 18.04, 64-bit
    • Windows, 64-bit
    • Android
  • MemryX Software Developer Hub (SDK)
  • CE / FCC Class A certification
  • Industrial -40°C to +85°C operating temperature range
  • 10% to 80% RH operating humidity, non-condensing

Resources

Paskelbta: 2025-01-08 | Atnaujinta: 2025-01-17