Organizations running machine learning and high performance computing workloads on Google's cloud computing platform now have a new option for accelerating application performance.
The company this week announced beta availability of Nvidia's Tesla V100 Graphics Processing Unit (GPU) for customers of the Google Compute Engine infrastructure as a service (IaaS) component and the Kubernetes Engine container orchestration system.
Organizations can take advantage of the new hardware accelerator to boost performance of large, very computationally intensive workloads. "Today’s most demanding workloads and industries require the fastest hardware accelerator," said Google product managers Chris Kleban and Ari Liberman, in a blog announcing availability of the new GPU.
The Nvidia Tesla V100 fits the bill by enabling up to 1 petaflop of hardware acceleration performance in certain high-end configurations, they said.
Nvidia describes its Tesla V100 GPU as the most advanced and powerful GPU it has built to date for accelerating artificial intelligence, graphics and HPC workloads. The GPU is based on Nvidia Volta, an architecture on which the company is building platforms for AI and other high- end applications. The Tesla V100 GPU is available in 32GB and 16GB configurations and offers performance equivalent to 100 CPUs in a single processor, according to Nvidia.
Google's goal in using the technology is to give customers the fastest available hardware accelerators currently available. Organizations can use up to 8 Tesla V100 GPUs, 96 virtual CPUs and 624GB of memory in a single virtual machine, the two Google product managers said.
Coupled with Nvidia Nvlink interconnect, organizations will be able to eventually get 300GBs of GPU bandwidth and performance boosts of up to 40 percent when running deep learning and HPC workloads, they noted.
The new Nvidia GPUs are currently available to Google cloud customers in the U.S and certain parts of Europe. Pricing for each V100 GPU starts at $2.48 per hour for on-demand virtual machines and 50 percent less, or $1.24 for pre-emptible VMs, which are a relatively lower-cost option for running large batch processing workloads. Per second billing is available as well for V100 enabled workloads.
Along with beta availability of the Tesla V100, Google also announced general availability of the Tesla P100, a GPU that the company has positioned as being ideal for CUDA (Compute Unified Device Architecture) parallel computing workloads.
The P100 offers organizations a balance between price and performance with configurations that support up to four GPUs, 96 virtual CPUs and 624 GB of memory per VM. Prices for P100-enabled systems start at $1.46 for on-demand VMs and $0.73 for pre-emptible VMs, according to Google.
As with all GPUs, Google's customers will be able to use Nvidia's Tesla V100 and P100 GPUs in a variety of custom configurations and storage options to meet their specific requirements and pay only for the resources they use, Kleban and Liberman said.