The short-list of true HPC cloud providers just got a little longer. Infrastructure-as-a-Service vendor SoftLayer has added high-end NVIDIA Tesla GPUs to its line of dedicated servers. According to the company, use cases include computation-intensive projects such as data mining, numerical and seismic analysis as well as video processing and 3D rendering.

As with all the company's offerings, the HPC servers will be available in a range of configurations. As a point of reference, for $879.00 per-month, users will get the entry level model, an Intel Xeon E5-2620 Sandy Bridge-based server outfitted with one GPU, 16GB of RAM and 500GB of storage.

SoftLayer, which opened for business in early 2006, claims to be the largest privately-held Infrastructure-as-a-Service provider in the world with over 100,000 servers located in datacenters spanning North America, Europe and Asia. While the company already has 25,000 customers ranging from technology startups to global corporations, the HPC offering should help extend its reach even further.

SoftLayer's product line can be categorized into three main solution types, dedicated, aka "bare metal," servers with an operating system on-board (this includes the new HPC server offering), dedicated servers running a hypervisor for customers who prefer to manage their own virtual machines, and the CloudLayer Service for customers who want virtual machines in the cloud.

Customers take care of self-provisioning and self-management. This means they use either online ordering or connect with a sales rep to order customized hardware or virtual machine solutions. For example, if the customer is ordering a dedicated server, they'll need to select the server configuration, the operating system, the type of hard drive, amount of RAM, and whether they want 10Gig or 1Gig server connections. There are a lot of options to choose from and the company says this helps distinguish it from the competition.

Of course, the big news here is that customers can now specify that they want a dedicated server with NVIDIA Tesla M2090 chips, which have the power to accelerate HPC workloads by up to 10x. SoftLayer's HPC offering comprises dual-processor Intel E5-2600 Sandy Bridge-based servers supporting one or two NVIDIA Tesla M2090 GPUs. This best-in-class NVIDIA GPU delivers up to 665 gigaflops of double precision performance, 1.3 teraflops of single precision performance, ECC memory error protection, and L1 and L2 caches.

On the CPU side, Intel's E5-2600 product family supports up to 8 cores per processor. Customers can select from the E5-2620, starting at $879, the E5-2650, starting at $1029, and the E5-2690, starting at $1179. Since this is an IaaS offering, support will naturally be limited to the network, the power, and the hardware, while software-level support will be up to the customers.

According to SoftLayer Chief Scientist Nathan Day, initial customer activity has come primarily from the oil and gas industry for companies that need to run seismic data workloads. The cloud provider is also seeing interest from the entertainment/media sector and content creators who use graphics processors for what they were originally designed, graphics processing. The science verticals tend to be the typical large HPC user, while on the entertainment and social media side, companies tend to be of the small-to-mid-size variety, or as Day puts it, "they are technically small businesses that are acting like like large enterprises in terms of their IT needs."

Day expects that the hybrid HPC offering will pull in a lot of customers who need the power of GPUs to run their workloads. To that end, the use case is any application that can take advantage of the GPU. Likely candidates will come from the scientific, number-crunching domain, or the graphics processing, video transcoding, and content creation arena, Day says.

When asked what sets apart their offering from others in the space, Day was quick to point out the benefits of a bare metal solution for HPC users. "So they don't have to pay the hypervisor tax," he comments. While Day credits Amazon Web Services with helping to popularize cloud computing, he notes that Amazon only offers virtual cycles; it does not provide the raw access to hardware of a dedicated solution.

Another plus for customers, Day says, is SoftLayer's wide variety of offerings and configurations. He also mentions the company's global reach. Datacenter facilities in Dallas, Seattle, Washington DC, Houston, San Jose, Amsterdam, and Singapore are all integrated into the SoftLayer network, which provides customers with over 2,000 Gbps of connectivity between 13 datacenters and 16 points of presence (PoPs). Additional PoPs include Hong Kong, Tokyo, Los Angeles, Denver, Chicago, New York, Atlanta, Miami, London and Frankfurt. The geographically-dispersed approach was designed to bring connectivity closer to the user, which brings us to our next topic, latency.

It's no surprise that HPC usually entails substantial data sets. Day says that it's common to see data sets on the order of hundreds of GBs, which need to be moved into the cloud for processing. While SoftLayer's robust network helps address latency concerns, it often makes more sense to transfer the data on a hard drive rather than deal with long upload times. For customers that are set up to perform deltas from the initial datasets, it's mainly the initial setup that presents a challenge. Regardless, Day says it's relatively easy to ship the data if necessary, which is exactly what their first HPC customer did. Subsequent customers are still doing calculations on how long the data transfer will take.

When asked the pivotal cloud question, why is it better to rent versus buy, Day responded that he expects to see a lot of project use initially as people start adapting to the model. "It's certainly easier to come to SoftLayer and get a fleet of servers with GPUs for a few months than it is that go purchase them outright, put them in a datacenter and pay for all the care and feeding that goes along with having servers in the datacenter," he says, adding the well-known cloud adage: "it basically takes what was a capital expense and makes it an operational expense to get the compute power to handle these workloads."