QNAP Systems - MUSTANG-F100-A10-R10

• Main FPGA: Intel® Arria® 10 GX1150 FPGA
• Operating Systems: PC:
• Ubuntu 16.04.3 LTS 64-bit, CentOS 7.4 64-bit, Windows 10 (More OS are coming soon)
• NAS: QTS (Installing “Mustang Card User Driver” in the QTS App Center is required.)
• Voltage Regulator and Power Supply: Intel® Enpirion® Power Solutions
• Memory: 8G on board DDR4
• Dataplane Interface: PCI Express x8 Compliant with PCI Express Specification V3.0
• Power Consumption (W): <60W
• Operating Temperature & Relative Humidity: 5°C~60°C (ambient temperature)，5% ~ 90%
• Cooling: Active fan: (50 x 50 x 10 mm) x 2
• Power Connector: Preserved PCIe 6-pin 12V external power
• Dip Switch/LED indicator: Up to 8 cards can be supported with operating systems other than QTS; QNAP TS- 2888X NAS supports up to 4 cards. Please assign a card ID number (from 0 to 7) to the Mustang-F100 by using rotary switch manually. The card ID number assigned here will be shown on the LED display of the card after power-up.

Intel® Vision Accelerator Design with Intel® Arria® 10 FPGA

As QNAP NAS evolves to support a wider range of applications (including surveillance, virtualization, and AI) you not only need more storage space on your NAS, but also require the NAS to have greater power to optimize targeted workloads. The Mustang-F100 is a PCIe-based accelerator card using the programmable Intel® Arria® 10 FPGA that provides the performance and versatility of FPGA acceleration. It can be installed in a PC or compatible QNAP NAS to boost performance as a perfect choice for AI deep learning inference workloads.

• Half-height, half-length, double-slot.
• Power-efficiency, low-latency.
• Supported OpenVINO™ toolkit, AI edge computing ready device.
• FPGAs can be optimized for different deep learning tasks.
• Intel® FPGAs supports multiple float-points and inference workloads.

OpenVINO™ toolkit

OpenVINO™ toolkit is based on convolutional neural networks (CNN), the toolkit extends workloads across Intel® hardware and maximizes performance.

It can optimize pre-trained deep learning model such as Caffe, MXNET, Tensorflow into IR binary file then execute the inference engine across Intel®-hardware heterogeneously such as CPU, GPU, Intel® Movidius™ Neural Compute Stick, and FPGA.

Get deep learning acceleration on Intel-based Server/PC

You can insert the Mustang-F100 into a PC/workstation running Linux® (Ubuntu®) to acquire computational acceleration for optimal application performance such as deep learning inference, video streaming, and data center. As an ideal acceleration solution for real-time AI inference, the Mustang-F100 can also work with Intel® OpenVINO™ toolkit to optimize inference workloads for image classification and computer vision.

• Operating Systems
  Ubuntu 16.04.3 LTS 64-bit, CentOS 7.4 64-bit, Windows 10 (More OS are coming soon)
• OpenVINO™ toolkit
  • Intel^® Deep Learning Deployment Toolkit
    • - Model Optimizer
    • - Inference Engine
  • Optimized computer vision libraries
  • Intel^® Media SDK
    *OpenCL™ graphics drivers and runtimes.
  • Current Supported Topologies: AlexNet, GoogleNet, Tiny Yolo, LeNet, SqueezeNet, VGG16, ResNet (more variants are coming soon)
  • Intel^® FPGA Deep Learning Acceleration Suite
• High flexibility, Mustang-F100-A10 develop on OpenVINO™ toolkit structure which allows trained data such as Caffe, TensorFlow, and MXNet to execute on it after convert to optimized IR.

QNAP NAS as an Inference Server

OpenVINO™ toolkit extends workloads across Intel® hardware (including accelerators) and maximizes performance. When used with QNAP’s OpenVINO™ Workflow Consolidation Tool, the Intel®-based QNAP NAS presents an ideal Inference Server that assists organizations in quickly building an inference system. Providing a model optimizer and inference engine, the OpenVINO™ toolkit is easy to use and flexible for high-performance, low-latency computer vision that improves deep learning inference. AI developers can deploy trained models on a QNAP NAS for inference, and install the Mustang-F100 to achieve optimal performance for running inference.