On Monday, Amazon Web Services announced that companies may run high-performance computing jobs on GPUs in its cloud.
The product, Cluster GPU Instance, will run within the Amazon Web Services (AWS) Elastic Compute Cloud (EC2), according to a blog post by Amazon.com chief technology officer Werner Vogels.
"We have already seen early customers out of the life sciences, financial, oil and gas, movie studios, and graphics industries becoming very excited about the power these instances give them," Vogels said on Monday. "Although everyone in the industry has known for years that general-purpose GPU processing is a direction with amazing potential, making major investments has been seen as high-risk given how fast moving the technology and programming was."
An instance will comprise of 22GB of memory, 33.5 EC2 Compute Units — equivalent to two Intel Xeon X5570 quad-core Nehalem-based processors — two Nvidia Tesla Fermi-architecture GPUs, 1,690GB of storage per instance, and 10Gb Ethernet for clustering.
The initial limit on the number of instances that can be clustered together is eight, though this may expand in the future, according to an AWS blog. Each instance can handle a teraflop of computation split across its two GPUs, giving a top level of 8 teraflops per cluster. The instances are Linux/Unix-based and do not support the Windows Server operating system.
The Cluster GPU Instance is the second clustered option that AWS has made available. It joins the CPU-based Cluster Compute instance, announced in July. Unlike other AWS instances, the CPU and GPU Cluster instances have access to 10Gb Ethernet and an HPC-bespoke networking architecture.
The combination of 10Gb Ethernet and a special networking architecture means that "all hosts in a cluster can communicate with other nodes in the cluster at the full capacity of the 10Gbps fabric at the same time without blocking. Clearly not all can communicate with a single member of the fleet at the same time, but the network can support all members of the cluster communicating at full bandwidth in unison. It's a sweet network, and it's the network that makes this a truly interesting HPC solution", AWS distinguished engineer James Hamilton wrote on his blog at the time of the CPU cluster launch.
Both Cluster GPU and Cluster Compute instances can be run within the same cluster, according to AWS, allowing customers to run workloads to their AWS cluster that require both heavy-CPU and GPU computations across a fast HPC-specific network.
Customers will need to use Tesla's Cuda instruction framework to get the most in-depth control over their instance, according to AWS. Customers will also be able to run workloads on the application via the industry standard OpenCL interface or via high-level languages such as Fortran, Python, C, C++ or Java.
Each Cluster GPU Instance is available for $2.10 (£1.30) an hour. It can currently be provisioned in the US via AWS's Northern Virginia Region. At the time of writing, AWS had not responded to a query about when the Clustered Instances will come to Europe.
The Nvidia Tesla GPU, as well as being utilised by AWS, is increasingly being used in high-performance computing. The world's current fastest supercomputer — the Chinese Tianhe-1A — uses 7,168 Tesla GPUs as part of its HPC architecture, along with 14,336 CPUs.