Amazon has announced the availability of Cluster GPU instances for high performance applications starting with an instance of 2 NVidia Tesla GPUs and going up to a cluster of 128 or more instances, appealing to financial analysis, imaging, biology, simulation, and other domains.
Amazon is offering a Cluster GPU instance for those who need a HPC platform but cannot afford one or do not want to support the initial cost of setting up one. The Cluster GPU instance belongs to the HPC group of instances, offering the equivalent of 33.5 EC2 Compute Units based on a pair of Intel Xeon X5570 quad-core Nehalem processors. A Compute Unit is the computing power offered by a 1 GHz Opteron or Xeon processor manufactured in 2007. Besides that, the GPU instance comes with a pair of NVIDIA Tesla Fermi M2050 GPUs, each unit offering: 515 double-precision Gflops or 1 single-precision Tflops, 3 GB DDR5 memory accessed on a 384-bit bus sustaining a bandwidth of 148 GB/sec. The instance also includes 22 GB of internal RAM and 1690 GB of disk storage. The user has the ability to organize 128 or more instances in clusters, communicating between them on 10 Gbps network links.
Currently only Linux is supported on a Cluster GPU instance, allowing developers to program in FORTRAN, C, C++, Java, Python, or Ruby against the GPU’s CUDA Driver API or CUDA C Runtime. Users can also make use of Amazon Elastic MapReduce web service and Public Data Sets when dealing with vast amounts of data which is common for HPC applications. The Cluster GPU instance is currently available only in US-N.Virginia region at $2.10/hour.
There are already other vendors offering GPU in the cloud such as Peer1, Nimbix, Penguin Computing or Hoopoe. They are providing solutions based either on NVidia Tesla or AMD FireStream technologies. The number of such vendors is supposed to rise in the following years because GPUs have become attractive for computational-heavy applications due to their high processing power and after including more and more general-purpose computation functionality taken from traditional CPUs.