GPU servers: Key questions and answers for companies 

Modern graphics cards contain concentrated computing power. For a long time now, they have not only been used to calculate video displays. In specialized servers for high-performance computing, they use their enormous computing power for applications such as artificial intelligence or simulations. Here are the key questions and answers regarding GPU servers.

Inhaltsverzeichnis

• What is a GPU server
• What is the difference between GPUs and CPUs?
• What are the advantages of GPU servers from the cloud?
• How can companies tell if HPC projects are better suited for GPUs or CPUs?
• What types of GPU servers does the Open Telekom Cloud offer?

When to comes to server performance, usually what is being referred to is the centrally installed processor, aka the CPU. However, there is another class of servers that can provide a dramatic increase in performance for complex calculations: GPU servers. Unlike traditional servers, GPU servers perform the majority of their computations in graphics processing units (GPUs). GPUs were originally developed to offload graphical calculations from a computer's main processor. Now, they are also important components for high-performance computing (HPC).

When companies install resources for high-performance computing, they have to decide what kind of processors the IT architecture should be based on: CPUs or GPUs?

PUs are the heart of every computer. They are designed as general-purpose tools that can flexibly handle a variety of different tasks, but they are not specialized in any of them. Like a Swiss Army knife with screwdrivers and scissors, they offer many possible applications, but should be replaced by a specialized tool when it comes to performing a specific task very often and very efficiently.

GPUs, on the other hand, are optimized to perform countless calculations in parallel, using an extremely high number of processor cores, so-called shaders. While a CPU can contain a maximum of 32 processor cores, a GPU can contain up to 5,000 shaders. However, the individual shader is significantly less powerful than a processor core of a CPU.

The high number of shaders enables GPUs to execute a multitude of similar operations in parallel. This enables them to process certain tasks quickly and cost-effectively. This not only predestines them for many graphics applications such as computer aided design (CAD), rendering, or graphics-intensive applications from the entertainment industry, but also qualifies them for many processes in the field of machine learning, e.g., for training artificial intelligence (AI), where large amounts of data often must be processed in the shortest possible time.

However, it cannot be generally stated that GPUs are better suited for high-performance computing than CPUs. The latter have their strengths where parallelization is not necessary or complex functionalities are to be expected.

Companies can use GPUs in high-performance computing to perform Big Data modeling and statistical analyses. The use of GPUs is particularly important in the areas of AI, machine learning and visualization.

However, most companies have highly fluctuating needs here, meaning they have to keep resources on hand which they are not using most of the time. This places a lot of strain on the IT department's budget. Not only that, but the hardware needs to be maintained and administered.

Cloud GPU servers remove the need to purchase expensive hardware that may only be used at full load for a few weeks a year, and which would otherwise lie idle.

They offer companies the opportunity to flexibly adapt available processing power to their needs, by simply booking and receiving immediately. The cloud provider takes care of all technical issues, administration, and maintenance. Another advantage is that it is not necessary to upgrade hardware as requirements increase, needing to replace them with newer and faster models.

If the task requires processing large amounts of image or sound data, this is an indication that GPUs are probably better suited. Examples include recognizing road signs and lane markings in autonomous driving or identifying tumors on images from an MRI. A look into the software solution requirements used can also help to identify the demands on the IT architecture. Ask an expert if you are unsure – the Open Telekom Cloud consultants are happy to help.

With us you can book Elastic Cloud Servers (ECS) – also called Virtual Machines – with GPU-optimized flavors. These contain a virtual graphics chip based on Nvidia T4 (pi2 and g6 flavors) and Nvidia V100 cards (p2, p2v and p2s flavors). These cards are particularly suitable for applications that process images and moving pictures, but also for cryptographic tasks.

Our GPU-accelerated ECS are divided into two series: the G-series and the P-series. The G series are graphics-accelerated servers suitable for 3D animation rendering and CAD. The P-series includes compute-accelerated or inference-accelerated machines suitable for deep learning, scientific computing, and CAE.

For full documentation on our GPU-accelerated Elastic Cloud Server, click here.

For performance-intensive computing operations with dedicated hardware, our bare metal servers physical.p1 and physical.p2 are available (Nvidia P100 and V100 GPU). These can be used for HPC scenarios or applications with increased data protection requirements, providing high GPU performance without usage of the hypervisor. In addition, both BMS flavors have an InfiniBand interface with up to 100 GBit/s to connect the BMS GPU servers in a cluster and use maximum bandwidth with low latency.

Full documentation of our GPU-accelerated bare metal servers can be found here.

Elastic Cloud Sever and Bare Metal Server are available in three pricing plans:

• Elastic – hourly billing
• Reserved – discounted monthly rates for continuous use
• ·Reserved Upfront – increased discounts through upfront payment

Here, the license costs for the operating system are priced according to the respective virtual machine.