Advances of GPUs

As mentioned in the previous blog, The GPSME project would help the SMEs have an easy access to the latest technical advancements of GPU through its toolkit without squandering resources. The toolkit would offer semi-automatic source code translation from CPU to GPU that would help in higher performance benefits. The project is based on cooperation of 4 companies and 2 universities and supported by REA-Research Executive Agency of EU. More information on this project can be found at

The intrinsically parallel GPU has always been a processor with ample computational resources and the capability of offering multi-thread processing. GPUs now provide better performance than CPUs due to their highly data parallel nature and the ability to achieve higher arithmetic intensity.

The major inhibiting factors on GPU use have previously been low on-board memory and poor double-precision performance. These have largely been overcome in the current generation of GPUs and GPU clusters, with the new generation of NVIDIA GPUs (codename Fermi) having an 8-fold improvement in performance at double precision, which further widens the performance gap between them and CPUs. Based on the Fermi architecture, the new NVIDIA Tesla 20-series offers off-the-shelf GPU cluster computing, delivering equivalent performance at 1/20th the power consumption and 1/10th the cost compared to the latest quad-core CPU.

Another significant factor is that GPU computational power has become inexpensive and widely available in many moderate computers with basic configurations (e.g. desktop PCs, laptops). The typical latest-generation card costs only a few hundreds euros, and these prices drop rapidly as new hardware emerges. Moreover, low-cost GPU clusters are now commercially available that provide exceptional computing power on the desktop. For example, the latest-generation NVIDIA Tesla S2050 4-GPU cluster retails at around €8,500.The parallel nature of the GPUs can provide vast speed gains for applications in which computational requirements are large and parallelism is substantial. Given their wide availability, GPUs are particularly suited to many SME related applications that target public users and require considerable level of heavy computation with limited resources and programming capacity. Even if the final target is a remote supercomputer, early testing and experimentation is often important before a major commitment is made to the use of high-performance computing facilities, which are normally expensive and require advance booking


Proudly working with: