Why?

Applied research in deep learning requires the fastest possible experiment turnaround times to rapidly explore multiple network architectures and manipulate and curate datasets to reduce solution delivery times for internal and external customers.

The Deep Learning Box is a system that is designed and built for this specific task.

Our goal is to build the fastest machine learning training device that you can plug and play for all your deep learning workloads.

Specs

Hardware

The key requirements of the highest deep learning throughput is constrained by the maximum power consumption while still adequately providing cooling without excessive noise. The choice of hardware relies on consumer grade products:

Professional Box

  • Motherboard: ASUS X99-E-10G
  • Memory: 64GB
  • CPU: Intel Core i7-6800K (6 Cores at 3.4 GHz)
  • GPUs: 4x NVIDIA GeForce GTX 1080Ti “Pascal” 11GB (44GB total)
  • Storage: 3x 3TB Raid5 (6TB available)
  • Power supply: 1600W

Developer Box

  • Motherboard: ASUS Prime X370-Pro
  • Memory: 32GB
  • CPU: AMD Ryzen 7 1800X (8 Cores at 3.6 GHz)
  • GPUs: 2x NVIDIA GeForce GTX 1080Ti “Pascal” 11GB (22GB total)
  • Storage: 2x 3TB Raid1 (3TB available)
  • Power supply: 750W

Software

The software stack is just as important as the hardware, allowing you to plug the Deep Learning Box into the power supply and immediately start using it for your workflow. We use:

  • OS: Ubuntu 16.04
  • CUDA Toolkit: 8.0
  • CuDNN: v5.1 and v6.0

Deep Learning Frameworks

We provide the following frameworks and high-level APIs:

  • TensorFlow
  • Caffe/Caffe2
  • Torch
  • PyTorch
  • Theano/libgpuarray
  • Keras
  • Lasagne
  • CNTK
  • Chainer
  • DIGITS
  • Paddle (NVIDIA Docker)
  • MXNet

Pricing

Professional Box

Description Price (without tax)
Deep Learning Pro Box per unit €9,999
Shipping Europe Free
Shipping USA €20

Developer Box

Description Price (without tax)
Deep Learning Dev Box per unit €4,999
Shipping Europe Free
Shipping USA €20

To order please complete the order form with information on shipping and preferred payment method. On completion we will send you the payment options. Shipping takes 6-7 working days after payment is received. For urgent orders Express Shipping may be requested which will take 3-4 working days.

Order Form Call Us: +49 1577 892 1979

All hardware components come with 3 year warranty.

Applications

Satellite and Aerial Imagery Analysis

Thanks to improved hardware and deep learning libraries, classification of satellite and aerial imagery can now be performed at astonishing speeds and at a much lower cost that was previously possible. Below we briefly describe the steps required for the classification of buildings using satellite imagery.

The analysis of building footprints via satellite imagery

This is made up of the following steps:

1 Building footprint detection. This is done using NVIDA’s DIGITS library (included in the Deep Learning Box Software Stack) which has shown to detect more than 90% of the large buildings such as commercial buildings. In regions where building boundaries are close together the detection of individual buildings can drop down to 70% or result in building not being clearly demarcated. However, depending on the region, detection thresholds may be modified to achieve better than 70% accuracies. See images below for examples of classified images:

2 Classification of buildings. This is done by training a model based on pre-classified data. Through sources such as OpenStreetMap and Local Government datasets we can obtain some pre-classified buildings for the study area. These are used to create classification models that can then be applied to new building footprints. The classification algorithms usually fall within an accuracy range of 95-70%. See image below showing pre-classified buildings.

Order Form Call Us: +49 1577 892 1979

About

We are a Berlin based consultancy specializing in Machine Learning and it’s applications in the Geospatial Industry. Our clients range from Australian Government, Caritas International to the German Auto Industry.