Deep Learning Workstation

Computing on cloud infrastructure is readily available with pre-constructed components and offers many operational advantages. The list of notable providers includes amongst others Amazon Web Services, Google Cloud Platform and Microsoft Azure. In many practical situations, it is nevertheless useful or even essential to have hardware on premises. This guide lists a few advantages of having a workstation on-premises, and it shall get you jumpstarted with ideas for what to buy in 2020, keeping in mind availability and pricing structure.

When to prefer a local workstation

Firstly, cost. Putting the benefit of scaling flexibility in cloud infrastructure aside, performing some serious deep-learning tasks can easily lead you into the four-figure range. In practice, this limits creativity because experiments are expensive. The break-even point for the profitability of a local workstation in comparison to cloud usage is reached rather rapid, usually after a few months.

Secondly, being connected to a computing instance via SSH or web portal is less comfortable for the daily workflow than being connected to your local workstation. Lags and connection drops are the daily reality for almost all users.

Thirdly, computing on cloud servers automatically means storing your data on the cloud provider's servers. Depending on the level of sensitivity of the data to be analyzed, this may be undesirable or even illegal. The unavailability of data due to such issues is a limiting factor for creativity and rapid prototyping.

Hardware Components

A workstation composed from the following list of components allows you to perform state-of-the-art research projects at a reasonable price in 2020. Note that assembly is an interesting part of the process. It will take an expert roughly 15 minutes, a hobbyist will need two hours. People without prior experience should set aside a weekend.

• Workstation case: As physical dimensions usually are no limiting factor in this setting, we recommend a workstation case that respects the E-ATX standard. Note that the chosen size standard affects the options for choice of main board. Good choice: Corsair Carbide Air 540.
• Power adaptor: The graphics card(s) will consume a lot of power. Depending on the number of cards to be installed, the power supply needs to be able to supply a sufficient amount of power. Otherwise, you'll be at risk of sudden power losses during work. If you want to be on the safe side with respect to future extensions, get a 1600W power supply. A higher quality certification will save a considerable amount of energy in daily use. Good choice: Super Flower Leadex 80 PLUS Platinum 1600W.
• Mainboard: Get a very reliable workstation-class mainboard without multimedia frills but with evenly spaced PCI slots so that all your graphics cards fit in well. Note the standards: X299 for Intel processors, X399 for AMD processors. Good choice: Asus WS X299 SAGE.
• CPU: Your choice: Intel or AMD. The advantage of AMD's processors is that they are cheaper and feature more threads. Intel's processors have a higher single-thread performance. Check out the new processors with VNNI. Note that Cascade Lake processors may require you to flash your main board's BIOS beforehands. Depending on the type of main board, this can be done without a CPU installed. Good choice: Intel Core i9-10920X.
• CPU cooler: A suitable cooling system to attach to your CPU. Good choice: Corsair Cooling Hydro Series H100i Platinum.
• RAM: Get 8x 16GB RAM DDR4. If you're running an Intel CPU, cheap RAM will suffice. Good choice: 8x G.SKILL RipJaws V 16GB DDR4-3200 CL16 DIMM.
• Hard drive(s): One 1TB NVMe SSD drive for operating system and scratch plus two large magnetic hard drives for storage. Good choice: 1x SSD 1TB Samsung M.2 PCI-E NVMe 970 EVO Plus, 2x Seagate BarraCuda SATA 6G, 7200 U/min, 2 TB.
• Graphics card(s): In 2020, we recommend one to four RTX 2080 Ti. These cards are being offered by a range of vendors, it doesn't much matter which one you choose. Adding a second card will not double the performance, but increase it roughly by half. To manage the heat inside the case, a blower-style fan version has to be selected if the workstation is to contain more than one card. Good choice: 1-4x ASUS GeForce RTX 2080 Ti Turbo 11G.
• Screen cables: Suitable choice of HDMI and DisplayPort cables, depending on the configuration and number of screens to be connected. Note that the ASUS card features 2x DisplayPort and 1x HDMI.
• Screen(s): One to three screens, each between 20 and 27 inches diagonal.
• Peripherals: Such as keyboard, mouse and USB hub.
• Operating system and software: Windows 10 and Linux are the candidates. Good choice: Download a recent version of Ubuntu Linux and boot it from a USB stick. Installation is a simple process. Afterwards, install all the required components from the Ubuntu's repositories and the pip package manager, such as compilers, interpreters, Python package managers, your IDE of choice, TensorFlow, Keras, PyTorch, cloud platform SDK and of course the Midnight Commander. For remote access, set up SSH for command line access and xRDP via SSH tunnel for remote graphical access.

The total cost of this system will vary between EUR 4000 and 7000 depending on the number of GPUs included. The assembly of the parts will pose no insurmountable obstacle to any experienced server administrator.