Cloud Provider Offers Up GPU Acceleration
Cloud infrastructure provider SoftLayer has updated their offering with hybrid servers. The company has now joined a small group of vendors offering GPU compute power on-demand. HPC in the Cloud’s Tiffany Trader covered the development in a feature story yesterday.
According to Trader’s report, the SoftLayer portfolio includes both dedicated and virtualized environments, but the hybrid servers are strictly bare metal. The ordering process includes customizations for storage, processor, OS, memory and network connectivity. Now that GPU-equipped nodes with the latest NVIDIA Tesla parts are available, users can get access to an additional 665 gigaflops of performance per server.
The hybrid systems will sport Intel’s new Sandy Bridge processors. With the low end server, which goes for $879 per month, you get a couple of Xeon E5-2620s, 16GB of memory, one GPU and 500GB storage memory. The host processor can be upgraded to the E5-2690, which pushes the rental to $1,179 per month.
SoftLayer owns more than 100,000 servers in Europe, Asia and North America. It’s likely the company has distributed a given number of GPUs per datacenter and will upgrade systems according to hybrid computing demand. Support is divided between SoftLayer and the user. Thee company handles all the hardware issues, with software, including provisioning, left to the customer.
Some early adopters of the cloud offering come from the oil and natural gas industry, who are using the infrastructure for seismic workloads. Also, media/entertainment companies are taking advantage of the GPUs to offload their graphics rendering. The company’s chief scientist, Nathan Day, expects the GPUs to attract research-based HPC users as well.
Of the cloud providers that offer GPU computing, Amazon is arguably the most popular player. Day drew a contrast between SoftLayer and Amazon’s offerings, explaining that Softlayer makes their hardware available through their bare metal service where Amazon only offers virtualization. When asked why his company offered bare metal services to users, Day responded, “So they don’t have to pay the hypervisor tax.”
To minimize latency, the company has 13 datacenters and 16 points of presence in major cities across the globe. By spreading out their infrastructure, SoftLayer aims to reduce ping times, physically locating their systems as close to the user as possible. While this might not fit for extremely low-latency applications like high speed trading, it doesn’t hurt either.
There are few HPC cloud vendors and even fewer offering GPU-equipped infrastructure. Companies that are considering running HPC applications with limited or non-existent in-house infrastructure or are thinking of migrating IT costs from capital to operational expenses, may find SoftLayer’s cloud a viable alternative.
Full story at HPC in the Cloud