The challenge
Training AI models can require substantial infrastructure resources, but the requirements are not consistent across every project. The architecture of the model, volume of training data and stage of development can all influence the compute, memory and storage needed to train it effectively.
This can make standard infrastructure configurations difficult to work with. A platform may provide more capacity than required in one area while limiting the resources that are most important to the workload, creating unnecessary cost or restricting performance.
Requirements can also change as an AI project progresses. Infrastructure that supports initial development and testing may need to expand as datasets grow, models become more complex or the project moves towards production.
Our approach
Bespoke hosting starts with the requirements of the AI workload. Our engineers work with customers to understand how the model is being developed and trained, the resources it currently requires and how those requirements are expected to change.
This assessment informs the design of the hosting environment. Compute, memory, storage and networking can be configured around the workload, rather than requiring the AI application to fit a predefined infrastructure package.
The appropriate hardware will depend on how the model operates. Different AI workloads can place different demands on processors, memory and storage, so these requirements need to be considered together when specifying the environment.
Future development is also factored into the infrastructure design. Where training requirements are expected to increase, the environment can be planned with a route to additional capacity, helping teams move from initial development and testing towards more demanding stages without starting again with a completely different hosting approach.
For organizations that do not want to manage the underlying infrastructure internally, the environment can be fully managed by our technical engineers, with monitoring and support available 24/7/365.
The outcome
Bespoke hosting gives AI teams an infrastructure environment designed around the requirements of their model and training process.
Resources can be allocated according to the areas where they are actually needed, while the environment can develop alongside the project as training requirements increase. This provides greater flexibility for AI workloads that do not fit comfortably within standard hosting configurations.
With the underlying infrastructure managed by Hyve, development teams can concentrate on training, testing and refining their models while retaining access to technical expertise as their hosting requirements evolve.
What should you consider when choosing hosting for AI model training?
Before selecting infrastructure for AI model training, it is important to understand the requirements of the workload itself. Key considerations include:
- How much compute is required for training?
- What are the memory requirements of the model?
- How much training data needs to be stored and processed?
- Does the workload have specific hardware requirements?
- How are resource requirements expected to change as the model develops?
- Will additional infrastructure be required for later development or production stages?
- Are there requirements around where the infrastructure and data are hosted?
- Does your team want to manage the underlying infrastructure internally or work with a managed hosting provider?
Answering these questions provides a clearer basis for designing infrastructure around the workload rather than selecting a hosting environment first and adapting the AI project to fit it.
Customer case study
Symbolic Mind needed a bespoke hosting environment to support the development and training of its EVA large language model (LLM). The team had a detailed understanding of its technical requirements, including substantial compute and memory requirements, but had struggled to find a provider able to translate them into a practical and cost-effective infrastructure solution.
Working with Symbolic Mind, Hyve designed a bespoke 56-core server for the initial development and testing stage. The infrastructure was planned as the first phase of a wider solution, giving the team a route to additional high-performance infrastructure as its training requirements increased.
Using the environment to train EVA, Symbolic Mind achieved training speeds of 3 GB per hour.
Vadim Asadov, VP at Symbolic Mind, said:
“We had a deep understanding of what we needed, and Hyve was able to translate this and build a solution that works for us.”
