Realization of an innovative system for complex calculations and pattern recognition in real time by using commercial graphics processors (GPU). Application in High Energy Physics experiments to select rare events and in medical imaging for CT, PET a

Research project


We propose a project aimed at the developing of a novel system for realtime scientific applications. Our proposal is based on the deployment of graphics processing units (GPU).

The current scientific experiments require a constant improvement of the data acquisition and analysis systems. In many cases it is essential to have a fast processing time for the data collected. This happens for example for big high energy physics experiments. In such experiments it is necessary to keep the large flux of data on buffers while selective and fast algorithms are performed to identify the very few events to store on disk. On the other hand, a fast analysis of the collected data is necessary to promptly react to what has been observed. For example, this is the case for the real time reconstruction of medical images during the diagnostics or the clinic therapy. Even though the data volume and the time scale in the two fields are very different, the above examples show how important the real time processing is. This is typically achieved developing systems dedicated to the specific application. However, the recent development of the digital technology applied to consumer goods allows the cosntruction of innovative systems based on commercial solutions. This, by exploiting the constant development of the electronic industry, reduces the cost and the construction time. Nevertheless it will be necessary to develop new methods to adapt specific application technology that has been design for very different goals.

The underlying idea of the project is to exploit the computational power of the modern graphics processor units for real time applications. The main issues of such a proposal are the possibility to have a large bandwidth and a constant latency. For data transmission we will study different solutions. These include the deployment of specific drivers with high bandwidth performances to be used with 10 and 40 Gb/s network cards, as well as the use of a FPGA to control the data transmission protocol. The most innovative aspect of this project is probably the solution of complex problems with the GPUs. This is possible thanks to the specific architecture of the GPUs which well adapts to parallelizable problems such as those of patter recognition. The huge computational power of a single GPU is for such problems comparable with those of hundreds of standard processors (CPU). Normal GPUs offer contained computational latency and can operate on data fluxes of few Gb/s. Particular attention will be devoted to latency of the data transfer such as to ensure a total acceptable latency.
Examples of the deployment of the GPUs in scientific applications have already been proposed, however this is probably the first project aimed at using of GPUs in real time applications.
The deployment of FPGA and GPUs in the same system will allow to take complex decisions in the first levels of trigger of HEP experiments up to event rates of tens of MHz, not even imaginable with the standard techniques. The possibility to effectively reduce the rate of collected data with an online selection with similar performances than those obtained offline will reduce the requests on the bandwidth. In addition, it will be easy to adapt the system also to the high level triggers allowing to reduce the cost of the current PC farms for the online selection.

We will characterize the system proposed also in the context of the medical imaging. Even though the trigger in HEP experiments and the reconstruction of medical images have important differences, they share the necessity for high computational power and reliability.
The final goals of the project are to demonstrate the effectiveness of the proposed solution in three fields different for acquisition rate, data volume, and computing intensity. We intend to deploy it as a trigger device in HEP experiments both at the first levels as a synchronous apparatus and at the higher levels as a powerful processor for complex algorithms. In medical diagnostics, we plan to use it to reconstruct images of Nuclear Magnetic Resonance (NMR), Positron Tomography, and Computed Tomography.

To evaluate the achievement of the research project we will benefit from the direct collaboration with the ATLAS and NA62 experiments at CERN, with the laboratory of the Nuclear Magnetic Resonance of the Rome University, with the UOC of “Tecniche Diagnostiche Avanzate” of the Rome Hospital Umberto I, with the PET system installed at the Physiology Institute of the CNR in Pisa, and department of Nuclear Medicine and of Surgical Anaesthetic and Radiological Sciences of the University Hospital in Ferrara.

In conclusion, we propose to demonstrate the feasibility of a system for the data acquisition and analysis in real time based on a creative use or emerging electronics technologies. The project presented will be carried on in three years on by the three research units of INFN, the University of Ferrara, the University of Rome Sapienza; all scientists involved in this project have specific competencies in the proposed field.
Effective start/end date1/1/12 → …




central processing units
data transmission
data acquisition
nuclear magnetic resonance
nuclear medicine
research projects
pattern recognition