SkyScale: On Demand Supercomputing Accelerating AI Workflows and Strategies

Tim Miller, President

As our readers are aware, artificial intelligence (AI) and deep learning are facilitating massive changes in consumer-facing industries, including automotive, fintech, and healthcare. But for many organizations aspiring to develop efficient deep learning platforms, unique infrastructural and technological requirements complicate their efforts. As CIO Applications analyzed cloud offerings making a difference in this regard, SkyScale emerged. SkyScale^® is San Diego based and provides a level of supercomputing to corporate AI efforts we’ve not seen elsewhere. SkyScale offers cloud-based, world-class multi-GPU hardware uniquely configured and tuned for HPC and AI workloads, leap-frogging key corporate implementation challenges and jump-starting the process. Jack Harrison, co-founder of SkyScale elaborated, “The super-fast parallel processing nature of GPU computing has been the primary driver behind accelerating AI and deep learning applications. SkyScale provides access to dedicated, multi-GPU hardware platforms in the cloud to companies desiring the fastest performance available as a service anywhere on the globe, accelerating not only their AI workflows, but also their AI strategies.”

There are significant barriers to entry to implement an AI strategy. In addition to employing data scientists, installing just a few multi-GPU nodes along with the requisite supporting equipment, facilities, power and cooling utilities and IT personnel can cost millions, putting it out of reach for most small—and many mid-sized—companies. Meanwhile, the cloud technology giants have focused primarily on CPU processing, only recently adding virtualized GPU capability. Regardless of the marketing spin, virtualization means shared equipment resulting in slower performance and multi-tenant security concerns.

Jack Harrison, Co-Founder

To test the relative performance of dedicated versus virtualized systems, CIO Applications secured data from benchmark tests executed by a 3rd party on 8-GPU instances of NVIDIA's^® latest generation Volta GPUs on SkyScale and on the industry’s largest cloud provider. In the comparative test, a standardized ResNet 50 deep neural network model for image classification was run on a Caffe2 deep learning framework. The dedicated SkyScale instance performed 30 percent faster at classifying images than the virtualized system hosted by the industry giant. A cost analysis then resulted in a combined 70 percent performance/cost benefit.

“Our on-demand dedicated offerings for project-based workflows enable remote access to supercomputing hardware when needed,” explains Tim Miller, President of SkyScale. “With these systems, users get root level control and don’t face the multi-tenant downsides related to virtualization. The resulting upside to our customers includes enhanced performance, flexibility, and multiple levels of cost savings based on usage and capital and facilities expense.”

SkyScale’s standard offerings include NVIDIA’s latest generation GPUs: the P100 (Pascal™) and the V100 (Volta™), as used in the benchmark tests. Both are available in single nodes of up to 16 GPUs, which can be clustered with Infiniband™ interconnects for the ultimate in AI training processing power. All of the elements—processors, GPUs, storage, and networking—are high-end components and tuned for AI workflows.

“We work with each customer to strategically create the best cloud infrastructure to meet their requirements,” adds Miller. “Once configured, the system rental price is quoted and fixed so there are no surprises, and direct access to technicians is a key differentiator.”

SkyScale’s systems are located in industry-leading HIPAA and ITAR compliant data centers on both coasts and in the south, ensuring complete physical and cyber security along with robust, redundant power, cooling, and networking. A tour of their San Diego data center is an education in itself.

During our investigation into cloud-based AI and ML technologies, SkyScale emerged as the leader in GPU supercomputing with uncompromising security and competitive pricing.