High Performance Computing (HPC) — Get a low-cost super computer by unleashing the power of GPUs
Articles,  Blog

High Performance Computing (HPC) — Get a low-cost super computer by unleashing the power of GPUs

High Performance Computing When you need to speed up your number crunching application by a factor of 100 or more, high performance computing is unbeatable. If your application – even when it deals with heavy numeric computation – takes too long for results, it’s ballast in your workflow. You will avoid using the product at all, or worse, won’t use it properly. That costs. Catalysts implements high performance computing based on a CPU-GPU system. GPUs – Graphic Processing Units – are many-core systems providing a highly parallel design. Perfect for many scientific simulations for example in the field of physics, chemistry, astrophysics, medicine, financial application or earth observation. GPU-based systems even start to conquer ordinary business applications. In heavy numeric computations, a GPU processing unit – also called streaming processor – is about 3 to 10 times slower compared to a CPU kernel. But there are so many of them that you can achieve incredible speed-ups. Nowadays, GPUs have several hundreds of these streaming processors. We at Catalysts have the algorithmic know-how to use this enormous, highly parallel potential for your heavy numeric calculations. Beside the enormous speed-up, many-core systems are leading in energy efficiency as well as cost efficiency. Your initial investment is about 20% lower for a CPU-GPU system than for a comparable CPU based system. Even energy costs for operation decrease by a factor of 10. Due to Moore’s Law, hybrid CPU-GPU-based systems can give a performance increase of about 1.000 times within one decade due to energy and cost efficient scalability. A future-oriented technology if you keep in mind that CPU based systems have a performance increase of about 100 folds within one decade. This table shows that the growth of GPU performance is likely to do so until 2017. Today, a NVIDIA GeForce GTX 580 reaches a theoretical peak performance of 1 TeraFLOP in a second for a price of 400 Euros. Extrapolated from today’s values, you will get 11 TeraFlop for 400 Euros in 2014. But what if you need 10 TeraFlop or more today? The answer is quite simple. We use multiple GPUs in one server or in the cloud. That is a scalable, anticipatory solution we adjust to your needs and requirements. To sum it up, your pros building on a GPU-CPU based system are incredible: You can tackle much bigger problems than you are currently thinking. You can choose better algorithms to solve your problems, that are faster, more accurate or that are giving you more detail. Your initial investment is lower compared to a CPU based system. With a GPU-CPU based system, you are building on a scalable, forward-looking technology that saves you energy costs as well. We at Catalysts have experience in setting up a multiple GPU system, using its advantage in our software products. You get a full range of IT solutions! With our knowledge, we are able to speed up your calculations. Imagine, instead of waiting hours for your results, it could only take you a few minutes or even seconds! We were already able to increase the performance of our consumer’s risk management tool by a factor of 400. Instead of waiting hours to receive the latest risk analysis result, they now get their results in a couple of minutes. We also have projects in the field of real-time image processing in our portfolio, built on a GPU. We unleash the power of GPU computing in the field of non-destructive analysis and imaging of the inner structure of materials, which is used for quality control or in the development of new materials. 28 pictures with a dimension ranging from 2.048 x 1.000 up to 8.192 x 1.000 are processed per second, including sophisticated algorithms like a fast Fourier transformation which is implemented in software. We are Catalysts. Software is our passion.


  • Scott Lane

    thats great. if you want to add a GPU using PCI express over cable there is quite a selection
    at maxexpansion. that combonation of software and hardware would be awesome.

  • Michael Aspetsberger

    Scott, indeed, if you want to go for a maximum gpu/accelerator configuration, you'll need to split up some of the PCI lanes, or convert PCIe3 to PCIe2.
    But it depends on the problem if one can really fully utilize all those devices while still staying below bandwidth limits.

Leave a Reply

Your email address will not be published. Required fields are marked *