Amazon Launches Supercomputing In The Cloud

Slideshow: Amazon's Case For Enterprise Cloud Computing

Amazon Web Services on Monday launched a new unit of computing to provide a form of supercomputing in its EC2 cloud for the masses. It's a form of parallel processing powered by clusters of graphics processors, the GPU Cluster Instance or graphics processing unit.

In the past, supercomputing has tended to be limited to academic and corporate researchers who had been given access to special-built high performance computing clusters. One of the requirements of the cluster was a high speed interconnect connecting the cluster nodes. The possibility of building such clusters and making them generally available came a step closer to reality with the emergence of 10 Gbs Ethernet switching equipment at affordable prices.

GPUs are designed to run on a cluster of EC2 servers, making use of the EC2 Cluster 10 Gbs network and achieving a level of high performance computing far above the capacity of EC2's simple Standard Instance. A "small" Standard Instance contains 1 Elastic Compute Unit of CPU power. An ECU is the equivalent of a 2007 Intel Xeon or AMD Opteron chip running at 1 to 1.2 GHz. The new GPU Cluster Instance contains 33.5 ECUs, along with 22 GBs of memory, versus 1 ECU and 1.8 GBs of memory with a small Standard Instance.

While the GPU Cluster Instance can be used for graphics rendering, it can also be used for other high performance computing tasks as well. It follows the simpler Cluster Compute Instance that AWS introduced earlier this year. GPU Cluster Instances are designed for customers "who need additional network and CPU performance for their large and complex HPC workloads," said Peter De Santis, general manager of EC2, in announcing the new instance. The GPU Cluster is intended to run jobs that can be subdivided into many smaller parallel processing jobs.

Slideshow: Amazon's Case For Enterprise Cloud Computing

The GPU Cluster Instance is also the Elastic Compute Cloud's biggest processing unit. It is priced at $2.10 an hour, compared to the Cluster Compute Instance's $1.60 an hour. A Small Standand Instance is 8.5 cents an hour, running Linux. The cluster instances are available to run Linux only; Windows is not being offered at this time.

High performance computing customers can use the Cluster GPUs to line up "massive parallel processing power," De Santis said in the announcement. Workloads that lend themselves to parallel processing could be speeded up by using the GPU Cluster Instances. Such workloads might include processing of masses of weather information or seismic data used in oil and gas exploration.

The new instance would be useful in rendering professional or scientific visualizations with masses of data behind them, doing large-scale floating point calculations or building simulations.

Having Cluster GPU computing available at an hourly charge may open a path to small and medium sized companies that in the past may have spurned such resources as too expensive to install themselves or too difficult to get on the waiting list of an academic or research high performance computing cluster system. "We are looking forward to seeing the innovation this will enable," De Santis added.

The Cluster GPU is based on NVIDIA Tesla M2050 processors, a general purpose processor derived from a graphics-oriented predecessor.

One early user is Calgary Scientific, a supplier of advanced medical imaging software. "For patients in critical care scenarios, every second cut from diagnosis to treatment can lead to a more positive outcome," said Pierre Lemire, CTO. The Amazon Cluster GPU instances "will help Calgary Scientific bring imagery from patients in need to the required medical professional with minimum infrastructure expense," he added.