The solution allows for more efficient use of already installed equipment
The Rubytech group presented an updated version of the Skala^r Artificial Intelligence Machine, adding the “Spektr AI” AI infrastructure management platform. The solution is designed for more efficient use of GPU accelerators: according to the company, their utilization can be increased from the usual 30–40% to 70–85%.
“Spektr AI” combines computing nodes into a single cluster and distributes resources among teams and tasks. The scheduler takes into account priorities and work order, and the ability to split GPUs into parts allows multiple workloads to run simultaneously. This makes it possible to use already installed equipment more intensively and postpone the purchase of additional accelerators as the number of tasks grows.
The platform supports launching and managing containers with AI applications and models, monitoring load, access control, and working with the corporate model catalog. This includes loading models from external and internal repositories.
“Spektr AI” runs on Deckhouse Kubernetes Platform, supports Red OS and Astra Linux. The cluster can be scaled up to 32 nodes, each equipped with eight GPUs. Resource and queue management is carried out through a single interface with role-based access control.
