Every CEO has data and analytics anxiety. GPUs are expensive. Analytics use of GPUs is very different during the various stages of the project such as model training and production. Many companies have GPUs sitting idle because they can’t keep them supplied with data. This is the definition of an opportunity for Service Providers to provide GPUs as a service for your customers.
Customers want to access a pool of GPU capacity that can be delivered to a host upon request. The service includes the ability to dynamically provision GPUs – attach and disconnect, add more or release GPUs, and scale according to the data requirements of each phase of the project. Simplicity is critical, so being able to do so in a fully automated manner leveraging RESTful APIs with as little as one line of code, saves time and money. Once the resources are no longer needed they are disconnected and available for use by another workload.
The 451 Group, in their Voice of the Service Provider: Differentiation, Vendor Selection & Budgets 2021 survey, found that “helping customers adapt to new processes and optimize infrastructure is important to how service providers create value.” When identifying capabilities critical to how their company creates value for its customers, service providers say that helping customers adapt to new processes and practices (e.g., DevOps) is most important (57%), followed by optimizing infrastructure or application performance (53%) and selecting/assembling the right technologies (52%).
Companies across the globe are beginning to incorporate AI, ML, and DL into standard business practices. Bob Hayes from Business over Broadway talks about survey results stating, “Businesses are leveraging the power of machine learning methods to help them extract better quality information, increase productivity, reduce costs and extract more value from their data. As the amount of data continues to grow along with the processing power of technology, businesses will continue to incorporate ML into their business.”
Researchers have found different AI / ML adoption rates. In one study, the adoption rate of ML Methods was 10%; in a 2020 study by McKinsey, the adoption rate of AI was 50%. Still, another study found that 42% of companies were currently using AI, and 40% of companies were planning on using AI in the next two years. Another 2020 study found that 59% of enterprises have machine learning initiatives either in production or at a proof-of-concept stage. These are new processes and practices, and SPs can deliver value to customers and differentiation in the marketplace.
Illustrating the depth of adoption, Bob Hayes looks at the difference in company size finding “that larger companies have higher adoption rates about ML methods. The largest enterprise companies (10,000+ employees) reported ML adoption rates of 61%. The smallest companies (0-49 employees) reported adoption rates of 33%. Of the smallest companies, a little over a quarter of them (27%) indicate that they are exploring the use of ML methods.” No matter what size customer the SP has, there is an opportunity to deliver value.
This is where the industry is today. The GPU market size is growing at 32% CAGR to an estimated size of $246 billion by 2028. GPUaaS (although the depth of reporting is light) is projected to be $15 billion by 2027. The pie is not getting smaller and more companies need resources, guidance, and access to the infrastructure that enables the extraction of greater value from their data.
So what are the key ingredients to building a service around GPU infrastructure?
- Resource agility – the ability to quickly and efficiently deliver infrastructure services that were not anticipated
- Ability to precisely match CPU resources with GPU resources
- A pool of resources that are unrestrained in the data center location (same rack or different racks fitting into the existing capacity planning, not dictating it)
- A pool of resources accessed by a common network – Ethernet
- Ability to add/remove resources as needed for many workloads
- Resources, once released, are immediately available for allocation to another workload without manual intervention
- Resources that do not add to the virtualization licensing costs of servers
- Complete automation of resource assignment
- The small initial investment with significant scale potential
- Ability to use existing infrastructure assets for deeper monetization
- Granular usage tracking
- A partner willing to build your service with you (GTM, engineering, service design)
We got you – Fungible composable infrastructure technologies enable a slow entry into GPUaaS with low risk and maximum return on investment. For information contact, [email protected], read the Fungible GPU-Connect (FGC) Blog, or watch the FGC Demo.